WO2016022583A2 - Write operations in a tree-based distributed file system - Google Patents

Write operations in a tree-based distributed file system Download PDF

Info

Publication number
WO2016022583A2
WO2016022583A2 PCT/US2015/043646 US2015043646W WO2016022583A2 WO 2016022583 A2 WO2016022583 A2 WO 2016022583A2 US 2015043646 W US2015043646 W US 2015043646W WO 2016022583 A2 WO2016022583 A2 WO 2016022583A2
Authority
WO
WIPO (PCT)
Prior art keywords
node
metadata
file system
tree
distributed
Prior art date
Application number
PCT/US2015/043646
Other languages
French (fr)
Other versions
WO2016022583A3 (en
Inventor
Mohit Aron
Ganesha Shanmuganathan
Original Assignee
Cohesity, Inc.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Cohesity, Inc. filed Critical Cohesity, Inc.
Publication of WO2016022583A2 publication Critical patent/WO2016022583A2/en
Publication of WO2016022583A3 publication Critical patent/WO2016022583A3/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/22Indexing; Data structures therefor; Storage structures
    • G06F16/2228Indexing structures
    • G06F16/2246Trees, e.g. B+trees
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00Error detection; Error correction; Monitoring
    • G06F11/07Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/14Error detection or correction of the data by redundancy in operation
    • G06F11/1402Saving, restoring, recovering or retrying
    • G06F11/1446Point-in-time backing up or restoration of persistent data
    • G06F11/1448Management of the data involved in backup or backup restore
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00Error detection; Error correction; Monitoring
    • G06F11/07Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/14Error detection or correction of the data by redundancy in operation
    • G06F11/1402Saving, restoring, recovering or retrying
    • G06F11/1446Point-in-time backing up or restoration of persistent data
    • G06F11/1448Management of the data involved in backup or backup restore
    • G06F11/1451Management of the data involved in backup or backup restore by selection of backup contents
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/21Design, administration or maintenance of databases
    • G06F16/219Managing data history or versioning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24Querying
    • G06F16/245Query processing
    • G06F16/2455Query execution
    • G06F16/24553Query execution of query operations
    • G06F16/24562Pointer or reference processing operations
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/907Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/907Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/908Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/95Retrieval from the web
    • G06F16/951Indexing; Web crawling techniques

Definitions

  • Embodiments presented herein generally relate to distributed storage and, more specifically, to write operations in a tree-based distributed file system.
  • Managing a file system generally requires managing a large amount of metadata about objects in the file system.
  • metadata includes information such as file sizes, directory structures, file permissions, physical storage locations of the files, and the like.
  • metadata includes information such as file sizes, directory structures, file permissions, physical storage locations of the files, and the like.
  • that metadata information is typically stored, it is often difficult to frequently and quickly clone the metadata for a file system.
  • a "flat" data store such as a NoSQL store (NoSQL stands for "Not Only Structured Query Language”).
  • NoSQL NoSQL stands for "Not Only Structured Query Language”
  • each item in the store can be accessed directly from a starting object (e.g., a root node).
  • a starting object e.g., a root node.
  • Quickly cloning the metadata stored in a flat store is difficult because each item in the store needs to be copied. Because the number of metadata entries can grow to be high (e.g., millions of entries), copying the metadata becomes very time-consuming, which prevents the file system from being quickly copied.
  • One embodiment includes a method for performing a write operation on a set of file system metadata.
  • the method includes receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes, and a view identifier that identifies a view, identifying a first root node in a file system metadata tree structure that is associated with the view, traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key, and modifying the value in a leaf node associated with the metadata key.
  • Each leaf node in the file system metadata tree stores a single metadata key-value pair.
  • Another embodiment includes a system with a processor and a memory.
  • the memory stores instructions that, when executed on the processor, perform an operation.
  • the operation itself includes receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes, and a view identifier that identifies a view, identifying a first root node in a file system metadata tree structure that is associated with the view, traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key, and modifying the value in a leaf node associated with the metadata key.
  • Each leaf node in the file system metadata tree store a single metadata key-value pair.
  • Another embodiment includes a non-transitory computer-readable medium storing instructions which, when executed by a processor, cause the processor to perform an operation.
  • the operation itself may include receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes and a view identifier that identifies a view, identifying a first root node in a file system metadata tree structure that is associated with the view, traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key, and modifying the value in a leaf node associated with the metadata key.
  • Each leaf node in the file system metadata tree stores a single metadata key-value pair.
  • Figure 1 illustrates a distributed storage system, according to one embodiment.
  • Figure 2 illustrates example file system metadata, according to one embodiment.
  • Figure 3A illustrates the file system metadata tree of Figure 2 after distributed storage system has taken a snapshot of the "current" view, according to one embodiment.
  • Figure 3B illustrates modifying, for the current view, metadata associated with a particular metadata key, after a snapshot has been taken, according to one embodiment.
  • Figure 4A illustrates adding a leaf node in a file system metadata tree that would cause the degree of a particular node to exceed t max , according to one embodiment.
  • Figure 4B illustrates deleting a leaf node in a file system metadata tree that would cause the degree of a particular node to fall below t min , according to one embodiment.
  • Figures 5A and 5B illustrate a read operation, and an associated sequence of acquiring and releasing synchronization locks, according to one embodiment.
  • Figures 5C and 5D illustrate a write operation, and an associated sequence of acquiring and releasing synchronization locks, according to one embodiment.
  • Figure 6 illustrates a cache system for caching metadata stored in a metadata tree structure, according to one embodiment.
  • Figure 7 illustrates an example server configured to perform the functionality of the storage appliance of Figure 1 , according to one embodiment of the present invention, according to one embodiment.
  • Figure 8 is a flow diagram of method steps for creating a snapshot of a current view, according to one embodiment.
  • Figure 9 is a flow diagram of method steps for reading, from a particular view, a value that is associated with a provided key, according to one embodiment.
  • Figure 10 is a flow diagram of method steps for writing, in a particular view, a value that is associated with a provided key, according to one embodiment.
  • Figure 1 1 is a flow diagram of method steps for determining whether a cache entry of a node stored in a local cache is valid, according to one embodiment.
  • Embodiments disclosed herein provide techniques for efficiently managing metadata in a distributed file system.
  • the techniques generally maintain a file system metadata tree structure that stores the file system metadata.
  • the tree structure includes a root node and leaf nodes, as well as intermediate nodes (for all but the smallest size trees).
  • the leaf nodes store file system metadata values for access.
  • This tree structure allows the file system metadata to be cloned by simply copying the root node and each pointer stored by the root node to the next level of the tree structure. Because the number of pointers for a root node is limited to a relatively small number, copying the root node does not require a large amount of time. Thus, clones of the file system metadata can be made quickly and frequently.
  • modifying the tree structure involves a maximum of log(N) write operations, where N is the number of leaf nodes in the tree. This number represents a manageable number of write operations, even when N is very large.
  • Embodiments disclosed herein also provide techniques for reading and writing values to the metadata tree in a synchronized manner with the use of a distributed lock system, as well as techniques for caching data in the metadata tree structure.
  • FIG. 1 illustrates a distributed storage system 100, according to one embodiment of the present invention.
  • the distributed storage system 100 includes storage appliances 106, which are coupled to a network 102.
  • Client systems 104 are also coupled to the network 102.
  • the network 102 is a computer network such as a local network or a global computer network like the Internet.
  • Storage appliances 106 are computer systems such as servers, that include processors, memory storing applications and data, and other computer components. As shown, storage appliances 106 each include a distributed lock authority 108, a distributed metadata manager 1 10, and a distributed node manager 1 1 1 . As described in detail below, the distributed metadata manager 1 10 and distributed node manager 1 1 1 operate together to manage stored file system metadata 1 14. Each storage appliance 106 includes a local cache 1 12 of nodes accessed by distributed metadata manager 1 10.
  • Distributed lock authority 108 stores and updates locks stored in lock storage 1 16.
  • the distributed metadata manager 1 10 maintains file system metadata in the form of nodes arranged in a tree (a "file system metadata tree structure").
  • Leaf nodes in the file system metadata tree structure include key- value pairs that associate metadata keys with values in the form of particular items of file system metadata.
  • the file system metadata is the metadata for a distributed file system and includes information such as file sizes, directory structures, file permissions, physical storage locations of the files, and the like.
  • a requestor requests a particular item of metadata by providing a metadata key to a distributed metadata manager 1 10, which traverses a file system metadata tree to find the item of metadata associated with that metadata key.
  • Embodiments are described herein using a tree storing file system metadata as a reference of example metadata stored in a tree.
  • the tree may be used to store metadata related to a single file.
  • leaf nodes of the tree store details about blocks of data within the individual file.
  • the distributed node manager 1 1 1 updates the nodes of the file system metadata tree structure in a distributed manner. More specifically, nodes in a file system metadata tree structure each have an associated node identifier (ID). The node ID acts as a node key for lookups performed by the distributed node manager 1 1 1 .
  • the nodes are stored in a distributed manner in that each node is stored on one or more storage appliances 106. Thus, when a distributed metadata manager 1 10 on a particular storage appliance 106 requests a node via the corresponding node key, that node may be stored in file system metadata 1 14 on that particular storage appliance 106 or within file system metadata 1 14 on a different storage appliance 106. In some embodiments, distributed node manager 1 1 1 stores the nodes in a distributed NoSQL store.
  • the distributed metadata manager 1 10 and the distributed node manager 1 1 1 operate at two different logical "levels" of information.
  • the first level is the level of the individual nodes, where distributed node manager 1 1 1 provides a node in response to a node key. This first level has no notion of a tree structure, and only "blindly" reads and writes nodes upon request, based on the node key.
  • the second level is the level of the tree, where distributed metadata manager 1 10 accesses nodes maintained by the distributed node manager 1 1 1 as a tree structure to provide file system metadata in response to metadata keys. Maintaining the file system metadata in a tree structure allows file system data to be managed efficiently, as is described in more detail below. Maintaining the individual nodes as a distributed NoSQL store allows the storage appliances 106 to maintain the file system metadata in a distributed manner.
  • the distributed lock authority 108 provides synchronization locks to nodes in the tree.
  • the distributed lock authority 108 provides for at least two types of locks: a read-lock and a write-lock.
  • a requesting entity such as a storage appliance 106, that holds a read-lock on a particular node is allowed to read the contents of that node but is not allowed to modify that node.
  • a requesting entity that holds a write- lock on a particular node is allowed to both read and modify the contents of that node.
  • Only one storage appliance 106 may obtain a write-lock on any particular node.
  • write-locks may also be thought of as exclusive locks.
  • More than one storage appliances 106 may hold a read-lock on any particular node.
  • read- locks may also be thought of as non-exclusive or shared locks.
  • Distributed lock authority 108 stores locks in lock storage 1 16.
  • Client system 104 is a computer system that includes processors, memory storing applications and data, and other hardware and software components. Client systems 104 may request file system metadata from storage appliances 106, via metadata keys. Although only two storage appliances 106 and two clients 104 are shown, distributed storage system 100 may include any number of storage appliances 106 and any number of clients 104.
  • FIG. 2 illustrates example file system metadata 200, according to one embodiment.
  • File system metadata 200 is stored in file system metadata storage 1 14 in one or more storage appliances 106.
  • the file system metadata 200 includes a file system metadata tree 201 (1 ) and a view table 220.
  • the file system metadata tree 201 (1 ) includes nodes. Each node includes a node identifier (ID) and a tree identifier (ID).
  • a tree may have three types of nodes: root nodes, such as the node with node ID Rootl , intermediate nodes, such the node with node ID Int1 , and leaf nodes, such as the node with node ID Datal .
  • Root nodes and intermediate nodes include pointers to other nodes.
  • node Rootl includes a pointer to node Int1 and a pointer to node Int2.
  • a leaf node is a terminal branch of the tree, i.e., a leaf node does not include pointers to other nodes.
  • the node IDs act as the node keys for distributed node manager 1 1 1 to perform a lookup operation.
  • the distributed metadata manager 1 10 can access the contents of a node by providing a node key (node ID) to the distributed node manager 1 1 1 .
  • node ID node key
  • the node IDs depicted in Figure 2 are represented as alphanumerical node ID's for clarity. However, node IDs may be in different forms, such as purely numerical, purely alphabetical, or other representation.
  • Root nodes are at the root of a file system metadata tree 201 and include pointers to other nodes. No node points to a root node. Intermediate nodes are nodes pointed to by at least one other node and include a pointer to at least one other node. Leaf nodes are nodes at the bottom of the tree. Leaf nodes are pointed to by at least one other node but that do not point to any other node. In one embodiment, e.g., one storing metadata for a file system, leaf nodes store key-value pairs of file system metadata. Metadata keys are lookup values by which a particular leaf node may be accessed, from distributed metadata manager 1 10. Values are the actual file system metadata, such as, for example, file size, physical storage location for file data, access permissions, and other information.
  • Each root node and each intermediate node stores pointers to other nodes in the file system metadata tree 201 .
  • the pointers are depicted in Figure 2 with arrows.
  • Each pointer includes a range of numerical and sorted metadata keys that can be found by following that particular pointer.
  • the Rootl node includes a pointer to an Int1 node and a pointer to an Int2 node.
  • the pointer to the Int1 node indicates a range of keys, specifically, metadata keys 3 through 6 can be found by following that pointer.
  • the pointer to the Int2 node indicates that metadata keys 9 through 12 can be found by following that pointer.
  • the Int1 and Int2 nodes also include pointers which indicate the metadata keys that can be found by following those pointers.
  • the leaf nodes of tree 201 (1 ) each include a metadata-key-value pair. Note, as a result, the pointer to each leaf node includes only the key of that leaf node.
  • each node has a corresponding "degree,” equal to the number of pointers from that particular node to other nodes.
  • the degree of a given node equals the number of direct children of that node.
  • a first node is a direct child of a second node if the second node has a pointer to the first node.
  • the degree of an intermediate node is prevented from dropping below a minimum threshold (also referred to herein as "t min ”) and from exceeding a maximum threshold (also referred to herein as "t max ").
  • the degree for a root node cannot exceed t max , although there is no minimum degree for the root node. In other words, a root node can have a number of direct children that is greater than or equal to zero and less than or equal to t max .
  • Distributed metadata manager 1 10 may add or delete nodes as more file system metadata is stored. If adding a node would cause the degree of a node to exceed t max for that node, then the distributed metadata manager 1 10 adds additional nodes to support the new node, as discussed in greater detail below with reference to Figure 4A. Similarly, if removing a node would cause the degree of a node to drop below t min for that node, then the distributed metadata manager 1 10 removes nodes, as discussed in greater detail below with reference to Figures 4B.
  • a file system metadata tree 201 may store file system data for one or more views. Views are logical "perspectives" of a file system relative to a particular point in time. In one example, a "current" view may represent a perspective of the file system that is up-to-date, and another view may represent a perspective of the file system that is 10 minutes old. Any particular file system metadata tree 201 may include information for a large number of different views.
  • the file system metadata 200 includes a view table 220 used to manage views.
  • the view table 220 stores data for accessing different views.
  • the view table 220 may includes a view identifier (ID) lookup 222 and a view entry lookup 226.
  • the view ID lookup 222 associates view handles with view IDs.
  • the view entry lookup 226 associates view IDs with view entries.
  • View entries include root nodes for three different view components: file metadata (labeled "fs"), object data (labeled "object”), and configuration data (labeled "config”).
  • File metadata stores information such as file sizes, pointers to physical storage for the file, and the like.
  • "Object" stores information to access file system data as objects instead of conventional files.
  • Each object typically includes the data itself, a variable amount of metadata, and other details.
  • "Config" stores configuration information for the file system.
  • Storage appliance 106 may access file system metadata, object data, or configuration data for a particular view by first retrieving the view ID from the view ID lookup 222 with the view handle, and then accessing the desired root node from the view entry lookup 226. Subsequently, the storage appliance 106 accesses that root node and traverses through the tree to access the particular leaf node, following pointers associated with the desired metadata key. Then, the storage appliance 106 reads the data in the leaf node to obtain the requested file system metadata.
  • TreelDs identify nodes as being associated with a particular view. As described in more detail below, distributed metadata manager 1 10 reads tree IDs to perform certain operations, such as modifying nodes, adding nodes, or deleting nodes. In Figure 2, only a single tree ID is depicted. However, a file system metadata tree 201 may include nodes that have different tree IDs.
  • Figure 3A illustrates the file system metadata tree 201 (1 ) of Figure 2 after distributed storage system 100 creates a snapshot of a "current" view. Creating a snapshot of a particular view effectively creates a clone of that view.
  • the distributed metadata manager 1 10 creates two copies of the root node associated with each view component of the current view. Each copy of the root node includes the pointers of the original root node. Further, each copy is given a new treelD. In some embodiments, the distributed metadata manager 1 10 deletes the original root node.
  • the distributed metadata manager 1 10 also updates the view table 220 so that the view handle associated with the current view points to one of the root node copies and the view handle associated with the view for the snapshot points to the other root node copy.
  • Providing two new root nodes, each with a different tree ID prevents updates made to a particular view (e.g., the current view) from affecting nodes of a different view (e.g., the snapshot view).
  • the snapshot may be associated with a particular time (2PM), indicating that the snapshot is a snapshot of the file system metadata taken at that particular time.
  • Root2 is associated with the current view, V F s-
  • distributed metadata manager 1 10 creates two copies of Rootl , each with a different tree ID.
  • Each of the copies of the root node (Root2 and Root3) includes pointers that point to the nodes that the original root node pointed.
  • Root2 includes a treelD of 2
  • Root3 includes a treelD of 3.
  • distributed metadata manager 1 10 modifies view table 220 so that the current view - V F s - is associated with Root2, and the snapshot view - VFS@2PM - is associated with Root3.
  • Distributed metadata manager 1 10 deletes Rootl because it is no longer needed.
  • Figure 3A only the file system metadata view component of each view is shown represented in the tree 201 (1 ) (i.e., there is no data associated with the object view component or the configuration component).
  • the distributed metadata manager 1 10 would create two copies of the root node for the file system view component, the object view component, and the configuration view component.
  • Figure 3B illustrates modifying metadata associated with a particular metadata key in the current view, after a snapshot has been created.
  • nodes with at least three different treelDs exist. Specifically, a treelD assigned to the immediately previous current view, a treelD assigned to the now- current view, and a treelD assigned to the snapshot exist.
  • a tree may include nodes with many more than three different tree ID's. That is, a tree may have many root nodes, each associated with different tree ID's.
  • the current view still accepts changes to metadata.
  • the metadata stored in leaf nodes pointed to by both the snapshot view and the current view is generally not changed by "new" data for the current view, because the snapshot represents a perspective of the file system metadata that is "frozen” in time.
  • the distributed metadata manager 1 10 copies that node.
  • the distributed metadata manager 1 10 may modify metadata associated with a particular metadata key (and with the current view) by traversing from the root node of the current view towards a "target node;" which is the leaf node that stores the value associated with the metadata key.
  • a target node which is the leaf node that stores the value associated with the metadata key.
  • pointers in each node includes information that indicates what metadata keys can be found by following that pointer.
  • the distributed metadata manager 1 10 traverses the metadata tree 201 towards the target node by repeatedly identifying and following the pointers that point towards the metadata key.
  • the distributed metadata manager 1 10 traverses the metadata tree 201 until reaching a node with a treelD different from the treelD associated with the current view.
  • the treelD associated with the current view is the treelD of the root node that is associated with the current view in the view table 220.
  • the distributed metadata manager 1 10 creates a copy of the node with the different treelD, and assigns a treelD to the copy that matches the treelD of the current view. That is, to modify the current view (e.g., to write a new value to a metadata key), the metadata manager 1 10 generates a new "branch" of the tree that is local to the current view and preserves an "old" branch of the tree associated with a different treelD (e.g., a treelD of another snapshot).
  • a copy of an intermediate node (now associated with the current view) has the same pointers as the intermediate node that was copied.
  • a copy of a leaf node has the same metadata-key-value pair as that leaf node.
  • the distributed metadata manager 1 10 modifies pointers of the parent of the copied node to point to the newly generated copy. Note, copies do not have the same NodelDs as the node being copied, so that the distributed node manager 1 1 1 is able to identify each node. That is, a node is a copy only at the logical level of the tree structure and not at the logical level associated with the distributed node manager 1 1 1 .
  • the distributed metadata manager 1 10 After creating the copied node and modifying the pointer of the parent node, the distributed metadata manager 1 10 continues traversing down the file system metadata tree 201 and compares the treelD of each encountered node with the treelD associated with the current view. Again, if the distributed metadata manager 1 10 encounters a treelD that is different from the treelD of the current view, the distributed metadata manager 1 10 copies that node and modifies the pointer of the parent node as described above. The distributed metadata manager 1 10 continues traversing the file system metadata tree 201 (copying nodes identified by the corresponding treelD as needed) until reaching the target node to be modified.
  • the metadata manager 1 10 creates a local copy of the leaf node for the current view if the treelD of the leaf node is different from the treelD of the current view.
  • the distributed metadata manager 1 10 modifies that leaf node, which includes the metadata key and the metadata to be modified.
  • the node to be modified for the current view is node Data3, which has a treelD of 1 .
  • distributed metadata manager 1 10 begins at the root node for the current view, which is Root2.
  • the distributed metadata manager 1 10 traverses the file system metadata tree 201 (1 ) towards node Data3. Since Data3 has a metadata key of 9, distributed metadata manager 1 10 follows the pointer to node Int2.
  • the distributed metadata manager 1 10 determines that the treelD for Int2 (1 ) is different from the treelD for Root2 (2), and therefore makes a copy of Int2, which is Int3.
  • Distributed metadata manager 1 10 also modifies the pointer from the parent of Int2 to instead point to the copy, Int3.
  • the copy, Int3, has the same pointers as Int2.
  • the distributed metadata manager 1 10 traverses further towards Data3, following the now-modified pointer from Root2 to Int3.
  • Int3 has the same treelD as Root2, distributed metadata manager 1 10 does not copy any nodes, and simply follows the pointer towards Data3.
  • the distributed metadata manager 1 10 examines Data3 and determines that the treelD of Data3 (1 ) is different from the treelD of Root2. Therefore, the distributed metadata manager 1 10 creates a copy of Data3, which is node Data5.
  • the copy - Data5 - includes the same metadata-key-value pair as Data3.
  • the distributed metadata manager 1 10 modifies the pointer of Int3 to point to Data5 instead of Data3.
  • distributed metadata manager 1 10 modifies the value in Data5 as requested. The modified value is indicated as Data3'.
  • Distributed metadata manager 1 10 may also add a value associated with a metadata key that is not currently in a file system metadata tree 201 .
  • the distributed metadata manager 1 10 traverses the file system metadata tree 201 towards a target node (here, a target node is an intermediate node that includes pointers that span a range of metadata keys include the metadata key corresponding to the value to be added), copying nodes that have a treelD that is different from the treelD of the root node associated with the current view.
  • the distributed metadata manager 1 10 continues to traverse the file system tree until arriving at the target node.
  • the target node may simply be the root node.
  • the distributed metadata manager 1 10 simply creates a new leaf node with the metadata-key-value pair as requested, and adds a pointer in the target node that points to the newly created leaf node.
  • the distributed metadata manager 1 10 traverses the file system metadata tree 201 , until arriving at a target node (here, a target node is an intermediate node that includes pointers that span a range of metadata keys include the metadata key corresponding to the value to be added), copying each node whose treelD is different from the treelD of the root node of the current view as described above (each node with a different treelD up to and including the leaf node is copied).
  • the distributed metadata manager 1 10 deletes the leaf node having the indicated metadata key and the same treelD as the root node, and removes the pointer from the parent of that leaf node to the leaf node.
  • root nodes and intermediate nodes have a maximum degree t max
  • intermediate nodes have a minimum degree t min . If adding or deleting a node would cause the degree of an intermediate node to exceed t max or drop below t min , or would cause the degree of a root node to exceed t max , then certain additional operations are performed, as described below.
  • Representing file system metadata as a file system metadata tree 201 provides several benefits. First, the process of creating a snapshot of a file system is 0(t max ), or in other words has a constant-order time complexity, since the maximum number of pointers to be copied is equal to t max . Thus, snapshots can be taken quickly and frequently.
  • modifying a node has 0(log(N)), where N is the number of leaf nodes, or in other words has logarithmic time complexity, since the number of nodes that are traversed to modify a node is equal to the tree depth, which is bounded by log(N).
  • modifying the file system metadata is a quick procedure.
  • Figure 4A illustrates adding a leaf node in a file system metadata tree 201 (2) that would cause the degree of a particular node to exceed t max .
  • distributed metadata manager 1 10 determines that a leaf node should be added, distributed metadata manager 1 10 traverses through the file system metadata tree 201 until arriving at the lowest intermediate node below which the metadata key for the node to be added would fit, and adds the node below that intermediate node, copying any nodes that have a treelD different from the treelD of the root for the current view along the way.
  • Distributed metadata manager 1 10 copies such nodes as described above.
  • the distributed metadata manager 1 10 While traversing through the file system metadata tree 201 , the distributed metadata manager 1 10 checks each encountered node to determine if that node has a degree that is equal to t max. If the degree of such a node is equal to t max , then the distributed metadata manager 1 10 splits that node. More specifically, the distributed metadata manager 1 10 creates a new intermediate node and transfers half of the pointers from the encountered node with the degree equal to t max to this new intermediate node. The distributed metadata manager 1 10 also adds a new pointer in the parent of the encountered node that points to the new intermediate node, and modifies the pointer within the parent node that points to the encountered
  • splitting intermediate nodes in this manner proactively conditions the file system metadata tree 201 so that intermediate nodes do not have too many children. If the degree of a node encountered while traversing through the file system metadata tree 201 is not equal to t max , then the distributed metadata manager 1 10 does not split the node. If a root node is split in this manner, then distributed metadata manager 1 10 creates a new root node that points to the two pieces of the split root node, effectively increasing the number of levels in the file system metadata tree 201 by 1 . If the distributed metadata manager 1 10 creates a new root node, then the distributed metadata manager 1 10 modifies the corresponding view entry in the view table 220 to point to the newly created root node.
  • the distributed metadata manager is adding a new leaf node, indicated as DataX+1 .
  • DataX+1 the node ID's, treelD's, and keys are not depicted in Figure 4A.
  • Distributed metadata manager 1 10 determines, based on the metadata key for DataX+1 , that DataX+1 belongs under intermediate node Int1 -1 . Because Int1 -1 already has a degree equal to t max , distributed metadata manager 1 10 splits Int1 -1 into Int1 -1 and Int1 -1 a.
  • distributed metadata manager 1 10 creates a new pointer in Root that points to Int1 - 1 a, and distributes half of the keys from Int1 -1 to Int1 -1 a, modifying the pointers for Int1 -1 and Int1 -1 a as needed.
  • Distributed metadata manager 1 10 determines that the degree of Root does not exceed t max , and therefore does not split the Root.
  • Figure 4B illustrates removing a leaf node in a file system metadata tree 201 (3) that would cause the degree of a particular node to fall below a minimum threshold, according to one embodiment.
  • the treelD of each node in Figure 4B is assumed to be identical.
  • the distributed metadata manager 1 10 While traversing the file system metadata tree 201 , the distributed metadata manager 1 10 checks each encountered node to determine if that node has a degree that is equal to t min . If the degree of such a node is equal to t min , then the distributed metadata manager 1 10 attempts to "shuttle" metadata keys from other nodes that are siblings of the encountered node to the encountered node. Two (or more) nodes are siblings when they share a direct parent node. Shuttling transfers metadata keys from one or more siblings of the encountered node such that the encountered node and the siblings each have a degree that is at least equal to t min .
  • distributed metadata manager 1 10 does not perform shuttling. Instead, distributed metadata manager 1 10 combines the encountered node with at least one sibling of the encountered node. More specifically, the distributed metadata manager 1 10 deletes either a sibling or the encountered node or the encountered node itself and stores keys from the deleted node in the remaining sibling. The distributed metadata manager 1 10 also modifies pointers of the parent of the encountered node to reflect the fact that a node has been deleted. If a node has a degree that is not equal to t min , then the distributed metadata manager 1 10 does not perform the steps described above for that node.
  • distributed metadata manager 1 10 deletes the root node and converts the direct child of the root node into the root node, with appropriate changes to the view table 220 as described above (i.e., modify the appropriate view entry to point to the new root node instead of the old root node).
  • Figure 4B illustrates an example of a leaf node being deleted from a file system metadata tree that would cause the degree of a particular node to fall below tmin, according to one embodiment.
  • the distributed metadata manager 1 10 processes a command to delete node DataX/2.
  • the treelD of each node in Figure 4B is assumed to be identical (in practice if not, nodes are copied as described above).
  • the distributed metadata manager 1 10 traverses the file system metadata tree 201 (3), arriving at Int1 .
  • the distributed metadata manager 1 10 determines that deleting DataX/2 would decrease the degree of Int1 below t min , and thus attempts to perform shuttling.
  • the distributed metadata manager 1 10 combines Int1 with sibling Int2. Thus, the distributed metadata manager 1 10 deletes Int2 and transfers metadata keys from Int2 to Int1 . Because Int1 is the only direct child of the root node, distributed metadata manager 1 10 deletes the root node and converts Int1 into the root node.
  • Distributed lock authority 108 accepts requests for locks on nodes in a file system metadata tree 201 . Before a distributed metadata manager 1 10 reads a particular node, the distributed metadata manager 1 10 obtains a read-lock for that node from a distributed lock authority 108. Similarly, before a distributed metadata manager 1 10 writes to a particular node, the distributed metadata manager 1 10 obtains a write-lock on that node from the distributed lock authority 108. When reading a leaf node, writing to a leaf node, adding a node, or deleting a node, the distributed metadata manager 1 10 obtains and releases locks from the distributed lock authority 108 in particular sequences.
  • Figure 5A illustrates a sequence of locks obtained for reading a leaf node in a file system metadata tree 201 (4), according to one embodiment.
  • the file system metadata tree 201 (4) includes a current view, associated with Rootl and a snapshot view, associated with Root2. Data has already been written to Data8 in the current view, and so copies of Int2, Int5, Int10, and Data8 exist with a treelD matching the treelD of Rootl (2). Root2 has a treelD of 3.
  • Figure 5B illustrates a sequence of states for selected nodes in the file system metadata tree 201 (4) associated with reading the leaf node. For clarity, Figure 5B does not depict all of the nodes shown in Figure 5A.
  • the leaf node to be read is associated with a particular metadata key.
  • the distributed metadata manager 1 10 To read this leaf node the distributed metadata manager 1 10 first obtains a read-lock on the root node associated with that view. The distributed metadata manager 1 10 maintains the read-lock on the root node until the entire read operation is completed (i.e., until the information is read from the leaf node). Then, the distributed metadata manager 1 10 follows the pointer in the root node that is associated with the range of metadata keys that includes the metadata key for the leaf node. The distributed metadata manager 1 10 obtains a read-lock on the next node and reads that node to determine the next node associated with a range of metadata keys that includes the metadata key for the leaf node.
  • the distributed metadata manager 1 10 obtains a read-lock on that next node and continues traversing the distributed metadata tree in this manner. Whenever the distributed metadata manager 1 10 has read-locks on three nodes, not counting the root node, the distributed metadata manager 1 10 releases the read-lock on the node that is highest up in the tree (except for the root node). Maintaining a read-lock on the root node in this manner prevents clone operations from proceeding for a particular root node while read-operations or write-operations with that root node. Since a clone operation would take a write-lock on the root node, while a clone operation is proceeding, no other operation (read or write) can proceed as all other operations hold at least a read-lock on the root node.
  • a distributed metadata manager 1 10 executes a command to read the value stored in the node Data3, for the current view.
  • the distributed metadata manager 1 10 begins with the root node associated with the current view, which is Rootl , and acquires a read-lock on that node.
  • the distributed metadata manager 1 10 examines the pointers of Rootl to determine that Int2' is the next node to read, and acquires a read-lock on that node.
  • the distributed metadata manager 1 10 examines the pointers of Int2' and determines that Int4 is the next node to read, and acquires a read-lock on that node.
  • the distributed metadata manager 1 10 then acquires a read-lock on Int8. Now having three read-locks on non-root nodes, the distributed metadata manager releases the read-lock on Int4. The distributed metadata manager acquires a read-lock on Data3, reads the value stored in Data3, and releases all read-locks.
  • Figure 5C illustrates a sequence of locks taken for writing to a leaf node in file system metadata tree 201 (4), according to one embodiment.
  • the file system metadata tree 201 (4) in Figure 5C includes additional nodes (illustrated with dashed lines) associated with a write-operation to be performed in node Data6, relative to the current view.
  • a distributed metadata manager 1 10 acquires a read-lock on the root node for that view. Then, the distributed metadata manager 1 10 follows a pointer in the root node towards the leaf node. The distributed metadata manager 1 10 obtains a read-lock on the next node and reads the pointers in the next node to determine which branch should be followed to reach the leaf node. The distributed metadata manager 1 10 continues traversing the tree in a manner similar to what is described above for the read-operation. The distributed metadata manager 1 10 maintains the read-lock on the root node as described above as well.
  • the distributed metadata manager 1 10 encounters a node with a treelD different from the treelD associated with the view being modified, then the distributed metadata manager 1 10 does not obtain another read-lock for the node at the next level. Instead, the distributed metadata manager 1 10 releases the read-lock on the parent of the encountered node. The distributed metadata manager 1 10 acquires a write-lock on the parent of the encountered node that has the treelD that is different than the treelD of the root node. Once the write-lock is acquired on the parent of the encountered node, the distributed metadata manager 1 10 determines whether the parent node has changed from when that parent node was released.
  • the distributed metadata manager 1 10 acquires a read-lock on the encountered node and creates a copy of the encountered node.
  • the distributed metadata manager 1 10 also modifies the pointer within the parent node that points to the encountered node to instead point to the copy of the encountered node.
  • the distributed metadata manager 1 10 releases the read-lock on the grandparent, and, starting from the copy of the encountered node, traverses down the file system metadata tree 201 as described above.
  • the distributed metadata manager 1 10 arrives at the leaf node to be modified, the distributed metadata manager modifies the value at that node (copying that node if warranted based on the treelD).
  • the distributed metadata manager 1 10 When the distributed metadata manager 1 10 acquires the write-lock on the parent node of the node that is encountered that has a treelD that is different than the root node, if the distributed metadata manager 1 10 determines that the parent node has changed, since the read-lock on the parent node was released, then the distributed metadata manager 1 10 simply restarts the whole procedure, beginning with acquiring a read lock on the root node and traversing into the tree. Distributed metadata manager 1 10 restarts this procedure because distributed metadata manager 1 10 does not have locks on a sufficient number of nodes. For example, assume that distributed metadata manager 1 10 holds a read-lock on nodes N, C, and GC, where N is a node, C is the child of N and GC is the grandchild of N.
  • distributed metadata manager 1 10 When the distributed metadata manager 1 10 inspects GC, distributed metadata manager 1 10 determines that GC should be modified. Therefore, distributed metadata manager 1 10 releases the locks on GC and C and take a write-lock on C. After obtaining this write-lock on C and reading the contents of C, distributed metadata manager 1 10 determines that C has changed in the time between giving up the read- lock and acquiring the write-lock and needs to be modified as well. In order to modify C, distributed metadata manager 1 10 might determine that distributed metadata manager 1 10 should obtain a write-lock on N and a read-lock on the parent of N (e.g., if N has a treelD different than the root node). However, since distributed metadata manager 1 10 may not hold a read-lock on the parent of N, distributed metadata manager 1 10 gives up all locks and restarts the entire procedure from the beginning.
  • a distributed metadata manager 1 10 executes a command to modify the value stored in Data6, for the current view.
  • the distributed metadata manager 1 10 obtains a read-lock on Rooti , a read-lock on Int2', a read-lock on Int5', and a read- lock on Int9.
  • the distributed metadata manager 1 10 determines that the treelD for Int9 (3) is different from the treelD for Int5' (2).
  • the distributed metadata manager 1 10 releases the read-lock on Int5' and Int9 and obtains a write-lock on Int5' and a read-lock on Int9.
  • the distributed metadata manager 1 10 creates a copy of Int9, which is Int9'.
  • the distributed metadata manager 1 10 modifies the pointer from Int5' to Int9 to instead point to the copy, Int9'.
  • the distributed metadata manager 1 10 obtains a read-lock on Int9' and releases the read lock on Int2'.
  • the distributed metadata manager 1 10 obtains a read-lock on Data6 and determines that the treelD of data6 (3) is different from the treelD of Rooti (2).
  • the distributed metadata manager 1 10 releases the read-locks on Int9' and Data6 and obtains a write-lock on Int9' and a read-lock on Data6.
  • the distributed metadata manager 1 10 creates a copy of Data6, which is Data6', and modifies the pointer of Int9' to instead point to Data6'. Then, the distributed metadata manager 1 10 obtains a write-lock on Data6', modifies the value stored in Data6', and releases all locks.
  • the distributed metadata manager 1 10 may also add a node or delete a node. Adding a node is similar to modifying the value in a leaf node. However, instead of traversing down to a leaf node, the distributed metadata manager 1 10 traverses down to an
  • intermediate node that would include a pointer to the leaf node, based on the metadata keys stored in the intermediate node, creates a new leaf node, and modifies the intermediate node to include a pointer to the newly added node.
  • Deleting a node is also similar to modifying the value in a leaf node. However, instead of traversing down to a leaf node, the distributed metadata manager 1 10 traverses to an intermediate node that includes the pointer to the node to be deleted, taking read-locks and write-locks as described above. If the intermediate node that includes the pointer to the node to be deleted is the only node that points to the node to be deleted, then the distributed metadata manager 1 10 deletes the node to be deleted. If the intermediate node that includes the pointer to the node to be deleted is not the only node that points to the node to be deleted, then the distributed metadata manager 1 10 does not delete that node. The distributed metadata manager 1 10 can determine whether the intermediate node is the only node that points to the node to be deleted by maintaining a count of the number of nodes that point to any particular node.
  • each storage appliance 106 stores a local cache 1 12 of nodes retrieved by distributed metadata manager 1 10. Because multiple storage appliances 106 may each access a local copy of a particular node, the storage appliances 106 need to determine whether cached data remains valid.
  • FIG. 6 illustrates a cache management system 600 that includes distributed lock authorities 108 within the storage appliances 106 and a local cache 1 12 stored on one of the storage appliances 106.
  • the storage appliance 106 When a storage appliance 106 first accesses a particular node, the storage appliance 106 stores that node in the local cache 1 12.
  • the storage appliance obtains a lock from the distributed lock authority 108.
  • the distributed lock authority 108 stores an indication (a lock entry 602) that the lock exists for a particular node and also stores a value referred to as a cache epoch 604 for that node.
  • the distributed lock authority 108 increments the cache epoch for the node by one.
  • the distributed metadata manager 1 10 When accessing a node, the distributed metadata manager 1 10
  • the distributed metadata manager 1 10 determines whether a copy stored in a local cache 1 12 is valid (if in the cache at all). To do so, the distributed metadata manager 1 10 obtains a read lock from the distributed lock authority 108 and determines the current cache epoch 604(1 ) for that node. If the cache epoch 604(1 ) read from the distributed lock authority 108 is equal to the cache epoch 604(2) stored in the local cache 1 12, then the copy of the node stored in the local cache 1 12 is valid. If the cache epochs do not match, then the copy of the node stored in the local cache is invalid. Note that even if distributed metadata manager 1 10 has a read-lock on a particular node, cache validity for that node is not guaranteed.
  • the distributed metadata manager 1 10 may have given the read-lock up at a prior time and then acquired a new read-lock on that node. In the period in which the distributed metadata manager 1 10 did not have a lock on the node, the node may have changed and thus the cached version of the node may no longer be valid.
  • FIG. 7 illustrates an example server 700 configured to perform the functionality of the storage appliance 106, according to one embodiment of the present invention.
  • the server 700 includes, without limitation, a central processing unit (CPU) 705, a network interface 715, a memory 720, and storage 730, each connected to a bus 717.
  • the server 700 may also include an I/O device interface 710 connecting I/O devices 712 (e.g., keyboard, display and mouse devices) to the server 700.
  • I/O device interface 710 connecting I/O devices 712 (e.g., keyboard, display and mouse devices) to the server 700.
  • the computing elements shown in server 700 may correspond to a physical computing system (e.g., a system in a data center) or may be a virtual computing instance executing within a computing cloud.
  • the CPU 705 retrieves and executes programming instructions stored in the memory 720 as well as stores and retrieves application data residing in the storage 730.
  • the interconnect 717 is used to transmit programming instructions and application data between the CPU 705, I/O devices interface 710, storage 730, network interface 715, and memory 720.
  • CPU 705 is included to be representative of a single CPU, multiple CPUs, a single CPU having multiple processing cores, and the like.
  • the memory 720 is generally included to be representative of a random access memory.
  • the storage 730 may be a disk drive storage device.
  • the storage 730 may be a combination of fixed and/or removable storage devices, such as fixed disc drives, removable memory cards, optical storage, network attached storage (NAS), or a storage area-network (SAN).
  • the memory 720 includes distributed lock authority 108, a distributed metadata manager 1 10, and a distributed node manager 1 1 1 .
  • Storage 730 includes file system metadata 1 14, local cache 1 12, and lock storage 1 16. Each element stored in memory 720 and storage 730 functions as described above.
  • Figure 8 is a flow diagram of method steps for creating a snapshot of a current view, according to one embodiment of the present invention. Although the method steps are described in conjunction with Figures 1 -7, persons skilled in the art will understand that any system configured to perform the method steps, in any order, falls within the scope of the present invention.
  • a method 800 begins at step 802, where distributed metadata manager 1 10 identifies the root node that is associated with the current view.
  • a view table 220 associates the current view with a root node as described above.
  • distributed metadata manager 1 10 identifies the set of nodes to which the root node points.
  • the distributed metadata manager 1 10 creates a second root node that is a copy of the first root node and therefore points to the same set of nodes that the first root node points to, and includes the same metadata keys as the first root node.
  • the distributed metadata manager 1 10 generates a third root node that is a copy of the first root node and therefore points to the same set of nodes that the first root node points to, and includes the same metadata keys as the first root node.
  • the distributed metadata manager 1 10 associates the second root node with the current view in the view table 220.
  • the distributed metadata manager 1 10 associates the third root node with the snapshot in the view table 220.
  • Figure 9 is a flow diagram of method steps for reading, from a particular view, a value that is associated with a provided metadata key, according to one embodiment of the present invention.
  • the method steps are described in conjunction with Figures 1 -7, persons skilled in the art will understand that any system configured to perform the method steps, in various alternative orders, falls within the scope of the present invention.
  • a method 900 begins at step 902, where the distributed metadata manager 1 10 identifies the root node that is associated with the view to be read from and acquires a read-lock on that root node.
  • the distributed metadata manager 1 10 identifies a node pointed to by the root node that lies on the path to the target node and sets that node as the current node.
  • the distributed metadata manager 1 10 acquires a read lock on the current node.
  • the distributed metadata manager 1 10 follows a pointer in the current node that points towards the provided metadata key and identifies the node at the end of the pointer as the "subsequent node.”
  • the distributed metadata manager 1 10 acquires a read lock on the subsequent node.
  • the distributed metadata manager 1 10 determines whether the subsequent node is the target node.
  • the target node is the node that is the leaf node associated with the provided metadata key. If the subsequent node is the target node, then the method proceeds to step 912.
  • the distributed metadata manager 1 10 reads data from the subsequent node and releases all read- locks. If, at step 910, the subsequent node is not the target node, then the method proceeds to step 913.
  • the distributed metadata manager 1 10 determines whether there are read locks acquired for at least three nodes other than the root node. If there are read locks acquired for at least three nodes other than the root node, then the method proceeds to step 914.
  • step 914 the distributed metadata manager 1 10 releases the read-lock for the highest node in the tree structure. After step 914, the method proceeds to step 916. If, in step 913, the distributed metadata manager 1 10 determines that read locks are acquired for less than three nodes other than the root node, then the method proceeds to step 916. At step 916, the distributed metadata manager 1 10 sets the subsequent node to the current node. After step 916, the method returns to step 906.
  • Figure 10 is a flow diagram of method steps for writing, in a particular view, a value that is associated with a provided metadata key, according to one embodiment of the present invention.
  • a method 1000 begins at step 1002, where distributed metadata manager 1 10 traverses the metadata tree, acquiring read locks, towards a target node. It should be understood that for a modify-value operation, the target node is the leaf node that stores that value.
  • the target node is the intermediate node to which distributed metadata manager 1 10 adds a leaf node as the new node.
  • the target node is the intermediate node whose direct child is the leaf node for deletion. While traversing, whenever the distributed metadata manager 1 10 has at least three read locks on nodes other than root nodes, the distributed metadata manager 1 10 releases the read lock that is highest up the tree.
  • the distributed metadata manager 1 10 determines whether a node is encountered that has a treelD different than that of the root node. If such a node is encountered, then the method proceeds to step 1004. If no such node is encountered, then the method proceeds to step 1012. At step 1004, the distributed metadata manager 1 10 releases the read lock on the encountered node and the parent of that node and acquires a write lock on the parent of the encountered node. At step 1006, the distributed metadata manager 1 10 determines whether the parent node is identical to before releasing the read lock. If the parent node is identical, then the method proceeds to step 1010. If the parent node is not identical, then the method proceeds to step 1008, at which the distributed metadata manager 1 10 releases all locks and restarts the write operation. After step 1008, the method returns to step 1002.
  • the distributed metadata manager acquires a read lock on the encountered node, copies the encountered node, and modifies the pointers in the parent to point to the copy of the encountered node instead of the encountered node.
  • the distributed metadata manager 1 10 determines whether the encountered node (either the one that was copied if the treelDs differed or simply an encountered node with the same treelD as the root if no node was encountered with a treelD different than the root node) is the target node. If the encountered node is not the target node, then the method returns to step 1002. If the encountered node is the target node, then the method proceeds to step 1014.
  • the distributed metadata manager 1 10 modifies the target node as specified in the write operation. More specifically, the distributed metadata manager 1 10 modifies, adds, or deletes a node, depending on what the write operation is.
  • Figure 1 1 is a flow diagram of method steps for determining whether a cache entry of a node stored in a local cache is valid, according to one embodiment of the present invention. Although the method steps are described in conjunction with Figures 1 -7, persons skilled in the art will understand that any system
  • a method 1 100 begins at step 1 102, where the distributed metadata manager 1 10 determines whether the distributed metadata manager 1 10 has a lock acquired for the node. If the distributed metadata manager 1 10 has a lock for the node, then the method proceeds to step 1 108. If, in step 1 102, the distributed metadata manager 1 10 does not have a lock for the node, then the method proceeds to step 1 106. At step 1 106, the distributed metadata manager 1 10 acquires a read- lock on the node and reads the cache epoch for that node from the distributed lock authority 108.
  • step 1 108 if the cache epoch stored in the distributed lock authority 108 is equal to the cache epoch stored in the local cache, then the method proceeds to step 1 104 and determines that the cache entry is valid. If, at step 1 108, the cache epoch is not equal to the cache epoch stored in the local cache, then the method proceeds to step 108, at which the distributed metadata manager 1 10 determines that the cache entry is invalid.
  • One advantage of the disclosed approach is that clones of the file system metadata can be made very quickly and frequently. Another advantage is that modification operations are relatively quick as well. A further advantage is that the file system metadata is stored in a distributed manner, which allows the metadata to be stored in a robust manner.
  • One embodiment of the invention may be implemented as a program product for use with a computer system.
  • the program(s) of the program product define functions of the embodiments (including the methods described herein) and can be contained on a variety of computer-readable storage media.
  • Illustrative computer-readable storage media include, but are not limited to: (i) non-writable storage media (e.g., read-only memory devices within a computer such as CD-ROM disks readable by a CD-ROM drive, flash memory, ROM chips or any type of solid- state non-volatile semiconductor memory) on which information is permanently stored; and (ii) writable storage media (e.g., floppy disks within a diskette drive or hard-disk drive or any type of solid-state random-access semiconductor memory) on which alterable information is stored.
  • non-writable storage media e.g., read-only memory devices within a computer such as CD-ROM disks readable by a CD-ROM drive, flash memory, ROM

Abstract

Techniques for cloning, writing to, and reading from file system metadata. Cloning involves identifying a first set of pointers included in a first root node in a file system metadata tree structure that stores file system metadata in leaf nodes of the tree structure, creating a first copy of the first root node that includes the first set of pointers, creating a second copy of the first root node that includes the first set of pointers, associating the first copy with a first view, and associating the second copy with a second view. Reading generally involves traversing the tree structure towards a target leaf node that contains data to be read. Writing generally involves traversing the tree structure in the same manner, but also creating copies of any nodes to be modified if those nodes are deemed to have a different treeID than a particular root node.

Description

WRITE OPERATIONS IN A TREE-BASED DISTRIBUTED FILE SYSTEM
BACKGROUND OF THE INVENTION
Field of the Invention
[0001] Embodiments presented herein generally relate to distributed storage and, more specifically, to write operations in a tree-based distributed file system.
Description of the Related Art
[0002] Managing a file system generally requires managing a large amount of metadata about objects in the file system. Such metadata includes information such as file sizes, directory structures, file permissions, physical storage locations of the files, and the like. In order to back-up file system information, it desirable to frequently and quickly "clone" or "snapshot" the metadata stored for a given file system. However, due to the manner in which that metadata information is typically stored, it is often difficult to frequently and quickly clone the metadata for a file system.
[0003] Frequently, such metadata is stored in a "flat" data store such as a NoSQL store (NoSQL stands for "Not Only Structured Query Language"). In such a "flat" store, each item in the store can be accessed directly from a starting object (e.g., a root node). Quickly cloning the metadata stored in a flat store is difficult because each item in the store needs to be copied. Because the number of metadata entries can grow to be high (e.g., millions of entries), copying the metadata becomes very time-consuming, which prevents the file system from being quickly copied.
[0004] One could avoid copying each node in a flat store by simply creating a root node copy that includes a reference to the original root node. Modifications to the metadata would then be made by creating new entries corresponding to those modifications, and updating the pointers from the root node copy to point to the new entries.
[0005] One issue with doing this, however, is that after many such cloning operations, the access time to nodes in the original flat store become very high, since this approach generates long chains of root nodes. Eventually, in order to reduce the access times, a coalescing operation can be performed, where each item in the original data store is copied each root node, so that each root node has a full set of metadata entries. However, because the number of metadata entries can be quite high, as described above, such coalescing operations result in at least some of the cloning operations requiring a large amount of time to complete.
SUMMARY OF THE INVENTION
[0006] One embodiment includes a method for performing a write operation on a set of file system metadata. The method includes receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes, and a view identifier that identifies a view, identifying a first root node in a file system metadata tree structure that is associated with the view, traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key, and modifying the value in a leaf node associated with the metadata key. Each leaf node in the file system metadata tree stores a single metadata key-value pair.
[0007] Another embodiment includes a system with a processor and a memory. The memory stores instructions that, when executed on the processor, perform an operation. The operation itself includes receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes, and a view identifier that identifies a view, identifying a first root node in a file system metadata tree structure that is associated with the view, traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key, and modifying the value in a leaf node associated with the metadata key. Each leaf node in the file system metadata tree store a single metadata key-value pair.
[0008] Another embodiment includes a non-transitory computer-readable medium storing instructions which, when executed by a processor, cause the processor to perform an operation. The operation itself may include receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes and a view identifier that identifies a view, identifying a first root node in a file system metadata tree structure that is associated with the view, traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key, and modifying the value in a leaf node associated with the metadata key. Each leaf node in the file system metadata tree stores a single metadata key-value pair.
BRIEF DESCRIPTION OF THE DRAWINGS
[0009] Figure 1 illustrates a distributed storage system, according to one embodiment.
[0010] Figure 2 illustrates example file system metadata, according to one embodiment.
[0011] Figure 3A illustrates the file system metadata tree of Figure 2 after distributed storage system has taken a snapshot of the "current" view, according to one embodiment.
[0012] Figure 3B illustrates modifying, for the current view, metadata associated with a particular metadata key, after a snapshot has been taken, according to one embodiment.
[0013] Figure 4A illustrates adding a leaf node in a file system metadata tree that would cause the degree of a particular node to exceed tmax, according to one embodiment.
[0014] Figure 4B illustrates deleting a leaf node in a file system metadata tree that would cause the degree of a particular node to fall below tmin, according to one embodiment.
[0015] Figures 5A and 5B illustrate a read operation, and an associated sequence of acquiring and releasing synchronization locks, according to one embodiment.
[0016] Figures 5C and 5D illustrate a write operation, and an associated sequence of acquiring and releasing synchronization locks, according to one embodiment. [0017] Figure 6 illustrates a cache system for caching metadata stored in a metadata tree structure, according to one embodiment.
[0018] Figure 7 illustrates an example server configured to perform the functionality of the storage appliance of Figure 1 , according to one embodiment of the present invention, according to one embodiment.
[0019] Figure 8 is a flow diagram of method steps for creating a snapshot of a current view, according to one embodiment.
[0020] Figure 9 is a flow diagram of method steps for reading, from a particular view, a value that is associated with a provided key, according to one embodiment.
[0021] Figure 10 is a flow diagram of method steps for writing, in a particular view, a value that is associated with a provided key, according to one embodiment.
[0022] Figure 1 1 is a flow diagram of method steps for determining whether a cache entry of a node stored in a local cache is valid, according to one embodiment.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0023] Embodiments disclosed herein provide techniques for efficiently managing metadata in a distributed file system. The techniques generally maintain a file system metadata tree structure that stores the file system metadata. In one embodiment, the tree structure includes a root node and leaf nodes, as well as intermediate nodes (for all but the smallest size trees). The leaf nodes store file system metadata values for access. This tree structure allows the file system metadata to be cloned by simply copying the root node and each pointer stored by the root node to the next level of the tree structure. Because the number of pointers for a root node is limited to a relatively small number, copying the root node does not require a large amount of time. Thus, clones of the file system metadata can be made quickly and frequently. Additionally, modifying the tree structure involves a maximum of log(N) write operations, where N is the number of leaf nodes in the tree. This number represents a manageable number of write operations, even when N is very large. Embodiments disclosed herein also provide techniques for reading and writing values to the metadata tree in a synchronized manner with the use of a distributed lock system, as well as techniques for caching data in the metadata tree structure.
[0024] Figure 1 illustrates a distributed storage system 100, according to one embodiment of the present invention. As shown, the distributed storage system 100 includes storage appliances 106, which are coupled to a network 102. Client systems 104 are also coupled to the network 102.
[0025] The network 102 is a computer network such as a local network or a global computer network like the Internet. Storage appliances 106 are computer systems such as servers, that include processors, memory storing applications and data, and other computer components. As shown, storage appliances 106 each include a distributed lock authority 108, a distributed metadata manager 1 10, and a distributed node manager 1 1 1 . As described in detail below, the distributed metadata manager 1 10 and distributed node manager 1 1 1 operate together to manage stored file system metadata 1 14. Each storage appliance 106 includes a local cache 1 12 of nodes accessed by distributed metadata manager 1 10.
Distributed lock authority 108 stores and updates locks stored in lock storage 1 16.
[0026] In one embodiment, the distributed metadata manager 1 10 maintains file system metadata in the form of nodes arranged in a tree (a "file system metadata tree structure"). Leaf nodes in the file system metadata tree structure include key- value pairs that associate metadata keys with values in the form of particular items of file system metadata. The file system metadata is the metadata for a distributed file system and includes information such as file sizes, directory structures, file permissions, physical storage locations of the files, and the like. A requestor requests a particular item of metadata by providing a metadata key to a distributed metadata manager 1 10, which traverses a file system metadata tree to find the item of metadata associated with that metadata key.
[0027] Embodiments are described herein using a tree storing file system metadata as a reference of example metadata stored in a tree. As another example, the tree may be used to store metadata related to a single file. In such a case, leaf nodes of the tree store details about blocks of data within the individual file. [0028] Of course, one of skill in the art will recognize that the tree structure disclosed herein may be adapted for use in a broad variety of situations and with a broad variety of data types.
[0029] The distributed node manager 1 1 1 updates the nodes of the file system metadata tree structure in a distributed manner. More specifically, nodes in a file system metadata tree structure each have an associated node identifier (ID). The node ID acts as a node key for lookups performed by the distributed node manager 1 1 1 . The nodes are stored in a distributed manner in that each node is stored on one or more storage appliances 106. Thus, when a distributed metadata manager 1 10 on a particular storage appliance 106 requests a node via the corresponding node key, that node may be stored in file system metadata 1 14 on that particular storage appliance 106 or within file system metadata 1 14 on a different storage appliance 106. In some embodiments, distributed node manager 1 1 1 stores the nodes in a distributed NoSQL store.
[0030] Note, the distributed metadata manager 1 10 and the distributed node manager 1 1 1 operate at two different logical "levels" of information. The first level is the level of the individual nodes, where distributed node manager 1 1 1 provides a node in response to a node key. This first level has no notion of a tree structure, and only "blindly" reads and writes nodes upon request, based on the node key. The second level is the level of the tree, where distributed metadata manager 1 10 accesses nodes maintained by the distributed node manager 1 1 1 as a tree structure to provide file system metadata in response to metadata keys. Maintaining the file system metadata in a tree structure allows file system data to be managed efficiently, as is described in more detail below. Maintaining the individual nodes as a distributed NoSQL store allows the storage appliances 106 to maintain the file system metadata in a distributed manner.
[0031 ] The distributed lock authority 108 provides synchronization locks to nodes in the tree. The distributed lock authority 108 provides for at least two types of locks: a read-lock and a write-lock. A requesting entity, such as a storage appliance 106, that holds a read-lock on a particular node is allowed to read the contents of that node but is not allowed to modify that node. A requesting entity that holds a write- lock on a particular node is allowed to both read and modify the contents of that node. Only one storage appliance 106 may obtain a write-lock on any particular node. Thus, write-locks may also be thought of as exclusive locks. More than one storage appliances 106 may hold a read-lock on any particular node. Thus, read- locks may also be thought of as non-exclusive or shared locks. Distributed lock authority 108 stores locks in lock storage 1 16.
[0032] Client system 104 is a computer system that includes processors, memory storing applications and data, and other hardware and software components. Client systems 104 may request file system metadata from storage appliances 106, via metadata keys. Although only two storage appliances 106 and two clients 104 are shown, distributed storage system 100 may include any number of storage appliances 106 and any number of clients 104.
[0033] Figure 2 illustrates example file system metadata 200, according to one embodiment. File system metadata 200 is stored in file system metadata storage 1 14 in one or more storage appliances 106. As shown, the file system metadata 200 includes a file system metadata tree 201 (1 ) and a view table 220. The file system metadata tree 201 (1 ) includes nodes. Each node includes a node identifier (ID) and a tree identifier (ID). Further, as shown, a tree may have three types of nodes: root nodes, such as the node with node ID Rootl , intermediate nodes, such the node with node ID Int1 , and leaf nodes, such as the node with node ID Datal . Root nodes and intermediate nodes include pointers to other nodes. For example, node Rootl includes a pointer to node Int1 and a pointer to node Int2. A leaf node is a terminal branch of the tree, i.e., a leaf node does not include pointers to other nodes.
[0034] The node IDs act as the node keys for distributed node manager 1 1 1 to perform a lookup operation. The distributed metadata manager 1 10 can access the contents of a node by providing a node key (node ID) to the distributed node manager 1 1 1 . Note, the node IDs depicted in Figure 2 are represented as alphanumerical node ID's for clarity. However, node IDs may be in different forms, such as purely numerical, purely alphabetical, or other representation.
[0035] Root nodes, as the name implies, are at the root of a file system metadata tree 201 and include pointers to other nodes. No node points to a root node. Intermediate nodes are nodes pointed to by at least one other node and include a pointer to at least one other node. Leaf nodes are nodes at the bottom of the tree. Leaf nodes are pointed to by at least one other node but that do not point to any other node. In one embodiment, e.g., one storing metadata for a file system, leaf nodes store key-value pairs of file system metadata. Metadata keys are lookup values by which a particular leaf node may be accessed, from distributed metadata manager 1 10. Values are the actual file system metadata, such as, for example, file size, physical storage location for file data, access permissions, and other information.
[0036] Each root node and each intermediate node stores pointers to other nodes in the file system metadata tree 201 . The pointers are depicted in Figure 2 with arrows. Each pointer includes a range of numerical and sorted metadata keys that can be found by following that particular pointer.
[0037] In the example file system metadata tree 201 (1 ) illustrated in Figure 2, the Rootl node includes a pointer to an Int1 node and a pointer to an Int2 node. The pointer to the Int1 node indicates a range of keys, specifically, metadata keys 3 through 6 can be found by following that pointer. The pointer to the Int2 node indicates that metadata keys 9 through 12 can be found by following that pointer. The Int1 and Int2 nodes also include pointers which indicate the metadata keys that can be found by following those pointers. As shown, the leaf nodes of tree 201 (1 ) each include a metadata-key-value pair. Note, as a result, the pointer to each leaf node includes only the key of that leaf node.
[0038] Also note, each node has a corresponding "degree," equal to the number of pointers from that particular node to other nodes. Stated differently, the degree of a given node equals the number of direct children of that node. A first node is a direct child of a second node if the second node has a pointer to the first node. In one embodiment, the degree of an intermediate node is prevented from dropping below a minimum threshold (also referred to herein as "tmin") and from exceeding a maximum threshold (also referred to herein as "tmax"). In some embodiments, tmax is a multiple of tmin. For example, in some embodiments, tmax = 3tmin or tmax = 2tmin. The degree for a root node cannot exceed tmax, although there is no minimum degree for the root node. In other words, a root node can have a number of direct children that is greater than or equal to zero and less than or equal to tmax.
[0039] Distributed metadata manager 1 10 may add or delete nodes as more file system metadata is stored. If adding a node would cause the degree of a node to exceed tmax for that node, then the distributed metadata manager 1 10 adds additional nodes to support the new node, as discussed in greater detail below with reference to Figure 4A. Similarly, if removing a node would cause the degree of a node to drop below tmin for that node, then the distributed metadata manager 1 10 removes nodes, as discussed in greater detail below with reference to Figures 4B.
[0040] A file system metadata tree 201 may store file system data for one or more views. Views are logical "perspectives" of a file system relative to a particular point in time. In one example, a "current" view may represent a perspective of the file system that is up-to-date, and another view may represent a perspective of the file system that is 10 minutes old. Any particular file system metadata tree 201 may include information for a large number of different views.
[0041] In one embodiment, the file system metadata 200 includes a view table 220 used to manage views. The view table 220 stores data for accessing different views. For example, the view table 220 may includes a view identifier (ID) lookup 222 and a view entry lookup 226. The view ID lookup 222 associates view handles with view IDs. The view entry lookup 226 associates view IDs with view entries. View entries include root nodes for three different view components: file metadata (labeled "fs"), object data (labeled "object"), and configuration data (labeled "config"). File metadata stores information such as file sizes, pointers to physical storage for the file, and the like. "Object" stores information to access file system data as objects instead of conventional files. Each object typically includes the data itself, a variable amount of metadata, and other details. "Config" stores configuration information for the file system. Storage appliance 106 may access file system metadata, object data, or configuration data for a particular view by first retrieving the view ID from the view ID lookup 222 with the view handle, and then accessing the desired root node from the view entry lookup 226. Subsequently, the storage appliance 106 accesses that root node and traverses through the tree to access the particular leaf node, following pointers associated with the desired metadata key. Then, the storage appliance 106 reads the data in the leaf node to obtain the requested file system metadata.
[0042] TreelDs identify nodes as being associated with a particular view. As described in more detail below, distributed metadata manager 1 10 reads tree IDs to perform certain operations, such as modifying nodes, adding nodes, or deleting nodes. In Figure 2, only a single tree ID is depicted. However, a file system metadata tree 201 may include nodes that have different tree IDs.
[0043] Figure 3A illustrates the file system metadata tree 201 (1 ) of Figure 2 after distributed storage system 100 creates a snapshot of a "current" view. Creating a snapshot of a particular view effectively creates a clone of that view. To create a snapshot of the current view, the distributed metadata manager 1 10 creates two copies of the root node associated with each view component of the current view. Each copy of the root node includes the pointers of the original root node. Further, each copy is given a new treelD. In some embodiments, the distributed metadata manager 1 10 deletes the original root node. The distributed metadata manager 1 10 also updates the view table 220 so that the view handle associated with the current view points to one of the root node copies and the view handle associated with the view for the snapshot points to the other root node copy. Providing two new root nodes, each with a different tree ID prevents updates made to a particular view (e.g., the current view) from affecting nodes of a different view (e.g., the snapshot view). As shown in Figure 3A, the snapshot may be associated with a particular time (2PM), indicating that the snapshot is a snapshot of the file system metadata taken at that particular time.
[0044] In the example file system metadata tree 201 (1 ) in Figure 3A, Root2 is associated with the current view, VFs- To create the snapshot of the file system data view component, having corresponding root node of Rootl , distributed metadata manager 1 10 creates two copies of Rootl , each with a different tree ID. Each of the copies of the root node (Root2 and Root3) includes pointers that point to the nodes that the original root node pointed. Also, Root2 includes a treelD of 2 and Root3 includes a treelD of 3. Additionally, distributed metadata manager 1 10 modifies view table 220 so that the current view - VFs - is associated with Root2, and the snapshot view - VFS@2PM - is associated with Root3. Distributed metadata manager 1 10 deletes Rootl because it is no longer needed. For simplicity, in Figure 3A only the file system metadata view component of each view is shown represented in the tree 201 (1 ) (i.e., there is no data associated with the object view component or the configuration component). However, to completely clone a view, the distributed metadata manager 1 10 would create two copies of the root node for the file system view component, the object view component, and the configuration view component.
[0045] Figure 3B illustrates modifying metadata associated with a particular metadata key in the current view, after a snapshot has been created. After creating a snapshot, nodes with at least three different treelDs exist. Specifically, a treelD assigned to the immediately previous current view, a treelD assigned to the now- current view, and a treelD assigned to the snapshot exist. Of course, after many snapshots are created, a tree may include nodes with many more than three different tree ID's. That is, a tree may have many root nodes, each associated with different tree ID's.
[0046] After a snapshot is made, the current view still accepts changes to metadata. However, because a snapshot exists, the metadata stored in leaf nodes pointed to by both the snapshot view and the current view is generally not changed by "new" data for the current view, because the snapshot represents a perspective of the file system metadata that is "frozen" in time. Thus, in order change file system metadata stored at a particular node for the current view, the distributed metadata manager 1 10 copies that node.
[0047] For example, the distributed metadata manager 1 10 may modify metadata associated with a particular metadata key (and with the current view) by traversing from the root node of the current view towards a "target node;" which is the leaf node that stores the value associated with the metadata key. As described above, pointers in each node includes information that indicates what metadata keys can be found by following that pointer. Thus, the distributed metadata manager 1 10 traverses the metadata tree 201 towards the target node by repeatedly identifying and following the pointers that point towards the metadata key. [0048] The distributed metadata manager 1 10 traverses the metadata tree 201 until reaching a node with a treelD different from the treelD associated with the current view. Note, the treelD associated with the current view is the treelD of the root node that is associated with the current view in the view table 220. At this point, the distributed metadata manager 1 10 creates a copy of the node with the different treelD, and assigns a treelD to the copy that matches the treelD of the current view. That is, to modify the current view (e.g., to write a new value to a metadata key), the metadata manager 1 10 generates a new "branch" of the tree that is local to the current view and preserves an "old" branch of the tree associated with a different treelD (e.g., a treelD of another snapshot). A copy of an intermediate node (now associated with the current view) has the same pointers as the intermediate node that was copied. A copy of a leaf node has the same metadata-key-value pair as that leaf node. After copying the node, the distributed metadata manager 1 10 modifies pointers of the parent of the copied node to point to the newly generated copy. Note, copies do not have the same NodelDs as the node being copied, so that the distributed node manager 1 1 1 is able to identify each node. That is, a node is a copy only at the logical level of the tree structure and not at the logical level associated with the distributed node manager 1 1 1 .
[0049] After creating the copied node and modifying the pointer of the parent node, the distributed metadata manager 1 10 continues traversing down the file system metadata tree 201 and compares the treelD of each encountered node with the treelD associated with the current view. Again, if the distributed metadata manager 1 10 encounters a treelD that is different from the treelD of the current view, the distributed metadata manager 1 10 copies that node and modifies the pointer of the parent node as described above. The distributed metadata manager 1 10 continues traversing the file system metadata tree 201 (copying nodes identified by the corresponding treelD as needed) until reaching the target node to be modified. Like the intermediate nodes, the metadata manager 1 10 creates a local copy of the leaf node for the current view if the treelD of the leaf node is different from the treelD of the current view. The distributed metadata manager 1 10 then modifies that leaf node, which includes the metadata key and the metadata to be modified. [0050] As shown in Figure 3B, the node to be modified for the current view is node Data3, which has a treelD of 1 . To modify this node, distributed metadata manager 1 10 begins at the root node for the current view, which is Root2. Next, the distributed metadata manager 1 10 traverses the file system metadata tree 201 (1 ) towards node Data3. Since Data3 has a metadata key of 9, distributed metadata manager 1 10 follows the pointer to node Int2. The distributed metadata manager 1 10 determines that the treelD for Int2 (1 ) is different from the treelD for Root2 (2), and therefore makes a copy of Int2, which is Int3.
[0051 ] Distributed metadata manager 1 10 also modifies the pointer from the parent of Int2 to instead point to the copy, Int3. The copy, Int3, has the same pointers as Int2. Next, the distributed metadata manager 1 10 traverses further towards Data3, following the now-modified pointer from Root2 to Int3. Because Int3 has the same treelD as Root2, distributed metadata manager 1 10 does not copy any nodes, and simply follows the pointer towards Data3. The distributed metadata manager 1 10 examines Data3 and determines that the treelD of Data3 (1 ) is different from the treelD of Root2. Therefore, the distributed metadata manager 1 10 creates a copy of Data3, which is node Data5. The copy - Data5 - includes the same metadata-key-value pair as Data3. Additionally, the distributed metadata manager 1 10 modifies the pointer of Int3 to point to Data5 instead of Data3. Finally, distributed metadata manager 1 10 modifies the value in Data5 as requested. The modified value is indicated as Data3'.
[0052] Distributed metadata manager 1 10 may also add a value associated with a metadata key that is not currently in a file system metadata tree 201 . To add such a value, the distributed metadata manager 1 10 traverses the file system metadata tree 201 towards a target node (here, a target node is an intermediate node that includes pointers that span a range of metadata keys include the metadata key corresponding to the value to be added), copying nodes that have a treelD that is different from the treelD of the root node associated with the current view. The distributed metadata manager 1 10 continues to traverse the file system tree until arriving at the target node. Of course, if the tree 201 is small, then the target node may simply be the root node. The distributed metadata manager 1 10 simply creates a new leaf node with the metadata-key-value pair as requested, and adds a pointer in the target node that points to the newly created leaf node.
[0053] To delete a value associated with a particular metadata key, the distributed metadata manager 1 10 traverses the file system metadata tree 201 , until arriving at a target node (here, a target node is an intermediate node that includes pointers that span a range of metadata keys include the metadata key corresponding to the value to be added), copying each node whose treelD is different from the treelD of the root node of the current view as described above (each node with a different treelD up to and including the leaf node is copied). The distributed metadata manager 1 10 deletes the leaf node having the indicated metadata key and the same treelD as the root node, and removes the pointer from the parent of that leaf node to the leaf node.
[0054] As described above, root nodes and intermediate nodes have a maximum degree tmax, and intermediate nodes have a minimum degree tmin. If adding or deleting a node would cause the degree of an intermediate node to exceed tmax or drop below tmin, or would cause the degree of a root node to exceed tmax, then certain additional operations are performed, as described below. Representing file system metadata as a file system metadata tree 201 provides several benefits. First, the process of creating a snapshot of a file system is 0(tmax), or in other words has a constant-order time complexity, since the maximum number of pointers to be copied is equal to tmax. Thus, snapshots can be taken quickly and frequently. Second, modifying a node has 0(log(N)), where N is the number of leaf nodes, or in other words has logarithmic time complexity, since the number of nodes that are traversed to modify a node is equal to the tree depth, which is bounded by log(N). Thus, modifying the file system metadata is a quick procedure.
[0055] Figure 4A illustrates adding a leaf node in a file system metadata tree 201 (2) that would cause the degree of a particular node to exceed tmax. As described above, when distributed metadata manager 1 10 determines that a leaf node should be added, distributed metadata manager 1 10 traverses through the file system metadata tree 201 until arriving at the lowest intermediate node below which the metadata key for the node to be added would fit, and adds the node below that intermediate node, copying any nodes that have a treelD different from the treelD of the root for the current view along the way. Distributed metadata manager 1 10 copies such nodes as described above.
[0056] While traversing through the file system metadata tree 201 , the distributed metadata manager 1 10 checks each encountered node to determine if that node has a degree that is equal to tmax. If the degree of such a node is equal to tmax, then the distributed metadata manager 1 10 splits that node. More specifically, the distributed metadata manager 1 10 creates a new intermediate node and transfers half of the pointers from the encountered node with the degree equal to tmax to this new intermediate node. The distributed metadata manager 1 10 also adds a new pointer in the parent of the encountered node that points to the new intermediate node, and modifies the pointer within the parent node that points to the encountered
intermediate node to correctly reflect the metadata keys now in the encountered intermediate node.
[0057] Splitting intermediate nodes in this manner proactively conditions the file system metadata tree 201 so that intermediate nodes do not have too many children. If the degree of a node encountered while traversing through the file system metadata tree 201 is not equal to tmax, then the distributed metadata manager 1 10 does not split the node. If a root node is split in this manner, then distributed metadata manager 1 10 creates a new root node that points to the two pieces of the split root node, effectively increasing the number of levels in the file system metadata tree 201 by 1 . If the distributed metadata manager 1 10 creates a new root node, then the distributed metadata manager 1 10 modifies the corresponding view entry in the view table 220 to point to the newly created root node.
[0058] In the example depicted in Figure 4A, the distributed metadata manager is adding a new leaf node, indicated as DataX+1 . (For clarity, the node ID's, treelD's, and keys are not depicted in Figure 4A). Distributed metadata manager 1 10 determines, based on the metadata key for DataX+1 , that DataX+1 belongs under intermediate node Int1 -1 . Because Int1 -1 already has a degree equal to tmax, distributed metadata manager 1 10 splits Int1 -1 into Int1 -1 and Int1 -1 a. Further, distributed metadata manager 1 10 creates a new pointer in Root that points to Int1 - 1 a, and distributes half of the keys from Int1 -1 to Int1 -1 a, modifying the pointers for Int1 -1 and Int1 -1 a as needed. Distributed metadata manager 1 10 determines that the degree of Root does not exceed tmax, and therefore does not split the Root.
[0059] Figure 4B illustrates removing a leaf node in a file system metadata tree 201 (3) that would cause the degree of a particular node to fall below a minimum threshold, according to one embodiment. For simplicity, the treelD of each node in Figure 4B is assumed to be identical. When distributed metadata manager 1 10 determines that a leaf node should be deleted, distributed metadata manager traverses through the file system metadata tree 201 until arriving at the intermediate node whose direct child is the node to be deleted, and deletes that direct child.
[0060] While traversing the file system metadata tree 201 , the distributed metadata manager 1 10 checks each encountered node to determine if that node has a degree that is equal to tmin. If the degree of such a node is equal to tmin, then the distributed metadata manager 1 10 attempts to "shuttle" metadata keys from other nodes that are siblings of the encountered node to the encountered node. Two (or more) nodes are siblings when they share a direct parent node. Shuttling transfers metadata keys from one or more siblings of the encountered node such that the encountered node and the siblings each have a degree that is at least equal to tmin.
[0061 ] If shuttling would still result in the encountered node (or any siblings) having a degree that is less than tmin, then distributed metadata manager 1 10 does not perform shuttling. Instead, distributed metadata manager 1 10 combines the encountered node with at least one sibling of the encountered node. More specifically, the distributed metadata manager 1 10 deletes either a sibling or the encountered node or the encountered node itself and stores keys from the deleted node in the remaining sibling. The distributed metadata manager 1 10 also modifies pointers of the parent of the encountered node to reflect the fact that a node has been deleted. If a node has a degree that is not equal to tmin, then the distributed metadata manager 1 10 does not perform the steps described above for that node.
[0062] If the above-described technique would cause the degree of a root node to equal 1 , and the direct child of that root node is not a leaf node, then distributed metadata manager 1 10 deletes the root node and converts the direct child of the root node into the root node, with appropriate changes to the view table 220 as described above (i.e., modify the appropriate view entry to point to the new root node instead of the old root node).
[0063] Figure 4B illustrates an example of a leaf node being deleted from a file system metadata tree that would cause the degree of a particular node to fall below tmin, according to one embodiment. In the example file system metadata tree 201 (3) illustrated in Figure 4B, the distributed metadata manager 1 10 processes a command to delete node DataX/2. For simplicity, the treelD of each node in Figure 4B is assumed to be identical (in practice if not, nodes are copied as described above). The distributed metadata manager 1 10 traverses the file system metadata tree 201 (3), arriving at Int1 . The distributed metadata manager 1 10 determines that deleting DataX/2 would decrease the degree of Int1 below tmin, and thus attempts to perform shuttling. Because Int2 has a degree equal to tmin, the distributed metadata manager 1 10 combines Int1 with sibling Int2. Thus, the distributed metadata manager 1 10 deletes Int2 and transfers metadata keys from Int2 to Int1 . Because Int1 is the only direct child of the root node, distributed metadata manager 1 10 deletes the root node and converts Int1 into the root node.
Traversing File System Metadata Trees Using Locks
[0064] Distributed lock authority 108 accepts requests for locks on nodes in a file system metadata tree 201 . Before a distributed metadata manager 1 10 reads a particular node, the distributed metadata manager 1 10 obtains a read-lock for that node from a distributed lock authority 108. Similarly, before a distributed metadata manager 1 10 writes to a particular node, the distributed metadata manager 1 10 obtains a write-lock on that node from the distributed lock authority 108. When reading a leaf node, writing to a leaf node, adding a node, or deleting a node, the distributed metadata manager 1 10 obtains and releases locks from the distributed lock authority 108 in particular sequences. Several operations on a file system metadata tree 201 , and the corresponding sequences of obtaining and releasing locks from a distributed lock authority 108, are now described with respect to Figures 5A-5D. [0065] Figure 5A illustrates a sequence of locks obtained for reading a leaf node in a file system metadata tree 201 (4), according to one embodiment. As shown, the file system metadata tree 201 (4) includes a current view, associated with Rootl and a snapshot view, associated with Root2. Data has already been written to Data8 in the current view, and so copies of Int2, Int5, Int10, and Data8 exist with a treelD matching the treelD of Rootl (2). Root2 has a treelD of 3. The rest of the file system metadata tree 201 (4) has treelD of 1 . For clarity, elements such as the metadata keys and node IDs are not depicted. Figure 5B illustrates a sequence of states for selected nodes in the file system metadata tree 201 (4) associated with reading the leaf node. For clarity, Figure 5B does not depict all of the nodes shown in Figure 5A.
[0066] Referring to Figures 5A and 5B together, the leaf node to be read is associated with a particular metadata key. To read this leaf node the distributed metadata manager 1 10 first obtains a read-lock on the root node associated with that view. The distributed metadata manager 1 10 maintains the read-lock on the root node until the entire read operation is completed (i.e., until the information is read from the leaf node). Then, the distributed metadata manager 1 10 follows the pointer in the root node that is associated with the range of metadata keys that includes the metadata key for the leaf node. The distributed metadata manager 1 10 obtains a read-lock on the next node and reads that node to determine the next node associated with a range of metadata keys that includes the metadata key for the leaf node. The distributed metadata manager 1 10 obtains a read-lock on that next node and continues traversing the distributed metadata tree in this manner. Whenever the distributed metadata manager 1 10 has read-locks on three nodes, not counting the root node, the distributed metadata manager 1 10 releases the read-lock on the node that is highest up in the tree (except for the root node). Maintaining a read-lock on the root node in this manner prevents clone operations from proceeding for a particular root node while read-operations or write-operations with that root node. Since a clone operation would take a write-lock on the root node, while a clone operation is proceeding, no other operation (read or write) can proceed as all other operations hold at least a read-lock on the root node. [0067] In the example file system metadata tree 201 (4) depicted in Figures 5A and 5B, a distributed metadata manager 1 10 executes a command to read the value stored in the node Data3, for the current view. The distributed metadata manager 1 10 begins with the root node associated with the current view, which is Rootl , and acquires a read-lock on that node. The distributed metadata manager 1 10 examines the pointers of Rootl to determine that Int2' is the next node to read, and acquires a read-lock on that node. The distributed metadata manager 1 10 examines the pointers of Int2' and determines that Int4 is the next node to read, and acquires a read-lock on that node. The distributed metadata manager 1 10 then acquires a read-lock on Int8. Now having three read-locks on non-root nodes, the distributed metadata manager releases the read-lock on Int4. The distributed metadata manager acquires a read-lock on Data3, reads the value stored in Data3, and releases all read-locks.
[0068] Figure 5C illustrates a sequence of locks taken for writing to a leaf node in file system metadata tree 201 (4), according to one embodiment. The file system metadata tree 201 (4) in Figure 5C includes additional nodes (illustrated with dashed lines) associated with a write-operation to be performed in node Data6, relative to the current view.
[0069] To write to a leaf node, a distributed metadata manager 1 10 acquires a read-lock on the root node for that view. Then, the distributed metadata manager 1 10 follows a pointer in the root node towards the leaf node. The distributed metadata manager 1 10 obtains a read-lock on the next node and reads the pointers in the next node to determine which branch should be followed to reach the leaf node. The distributed metadata manager 1 10 continues traversing the tree in a manner similar to what is described above for the read-operation. The distributed metadata manager 1 10 maintains the read-lock on the root node as described above as well.
[0070] However, if the distributed metadata manager 1 10 encounters a node with a treelD different from the treelD associated with the view being modified, then the distributed metadata manager 1 10 does not obtain another read-lock for the node at the next level. Instead, the distributed metadata manager 1 10 releases the read-lock on the parent of the encountered node. The distributed metadata manager 1 10 acquires a write-lock on the parent of the encountered node that has the treelD that is different than the treelD of the root node. Once the write-lock is acquired on the parent of the encountered node, the distributed metadata manager 1 10 determines whether the parent node has changed from when that parent node was released.
[0071 ] If the parent node has not changed, then the distributed metadata manager 1 10 acquires a read-lock on the encountered node and creates a copy of the encountered node. The distributed metadata manager 1 10 also modifies the pointer within the parent node that points to the encountered node to instead point to the copy of the encountered node. At this point, the distributed metadata manager 1 10 releases the read-lock on the grandparent, and, starting from the copy of the encountered node, traverses down the file system metadata tree 201 as described above. When the distributed metadata manager 1 10 arrives at the leaf node to be modified, the distributed metadata manager modifies the value at that node (copying that node if warranted based on the treelD).
[0072] When the distributed metadata manager 1 10 acquires the write-lock on the parent node of the node that is encountered that has a treelD that is different than the root node, if the distributed metadata manager 1 10 determines that the parent node has changed, since the read-lock on the parent node was released, then the distributed metadata manager 1 10 simply restarts the whole procedure, beginning with acquiring a read lock on the root node and traversing into the tree. Distributed metadata manager 1 10 restarts this procedure because distributed metadata manager 1 10 does not have locks on a sufficient number of nodes. For example, assume that distributed metadata manager 1 10 holds a read-lock on nodes N, C, and GC, where N is a node, C is the child of N and GC is the grandchild of N. When the distributed metadata manager 1 10 inspects GC, distributed metadata manager 1 10 determines that GC should be modified. Therefore, distributed metadata manager 1 10 releases the locks on GC and C and take a write-lock on C. After obtaining this write-lock on C and reading the contents of C, distributed metadata manager 1 10 determines that C has changed in the time between giving up the read- lock and acquiring the write-lock and needs to be modified as well. In order to modify C, distributed metadata manager 1 10 might determine that distributed metadata manager 1 10 should obtain a write-lock on N and a read-lock on the parent of N (e.g., if N has a treelD different than the root node). However, since distributed metadata manager 1 10 may not hold a read-lock on the parent of N, distributed metadata manager 1 10 gives up all locks and restarts the entire procedure from the beginning.
[0073] In the example file system metadata tree 201 (4) depicted in Figures 5C and 5D, a distributed metadata manager 1 10 executes a command to modify the value stored in Data6, for the current view. The distributed metadata manager 1 10 obtains a read-lock on Rooti , a read-lock on Int2', a read-lock on Int5', and a read- lock on Int9. The distributed metadata manager 1 10 determines that the treelD for Int9 (3) is different from the treelD for Int5' (2). The distributed metadata manager 1 10 releases the read-lock on Int5' and Int9 and obtains a write-lock on Int5' and a read-lock on Int9. Then the distributed metadata manager 1 10 creates a copy of Int9, which is Int9'. The distributed metadata manager 1 10 modifies the pointer from Int5' to Int9 to instead point to the copy, Int9'. The distributed metadata manager 1 10 obtains a read-lock on Int9' and releases the read lock on Int2'. The distributed metadata manager 1 10 obtains a read-lock on Data6 and determines that the treelD of data6 (3) is different from the treelD of Rooti (2). Thus, the distributed metadata manager 1 10 releases the read-locks on Int9' and Data6 and obtains a write-lock on Int9' and a read-lock on Data6. The distributed metadata manager 1 10 creates a copy of Data6, which is Data6', and modifies the pointer of Int9' to instead point to Data6'. Then, the distributed metadata manager 1 10 obtains a write-lock on Data6', modifies the value stored in Data6', and releases all locks.
[0074] In addition to modifying the value stored in a leaf node, the distributed metadata manager 1 10 may also add a node or delete a node. Adding a node is similar to modifying the value in a leaf node. However, instead of traversing down to a leaf node, the distributed metadata manager 1 10 traverses down to an
intermediate node that would include a pointer to the leaf node, based on the metadata keys stored in the intermediate node, creates a new leaf node, and modifies the intermediate node to include a pointer to the newly added node.
Deleting a node is also similar to modifying the value in a leaf node. However, instead of traversing down to a leaf node, the distributed metadata manager 1 10 traverses to an intermediate node that includes the pointer to the node to be deleted, taking read-locks and write-locks as described above. If the intermediate node that includes the pointer to the node to be deleted is the only node that points to the node to be deleted, then the distributed metadata manager 1 10 deletes the node to be deleted. If the intermediate node that includes the pointer to the node to be deleted is not the only node that points to the node to be deleted, then the distributed metadata manager 1 10 does not delete that node. The distributed metadata manager 1 10 can determine whether the intermediate node is the only node that points to the node to be deleted by maintaining a count of the number of nodes that point to any particular node.
Caching System
[0075] In one embodiment, each storage appliance 106 stores a local cache 1 12 of nodes retrieved by distributed metadata manager 1 10. Because multiple storage appliances 106 may each access a local copy of a particular node, the storage appliances 106 need to determine whether cached data remains valid.
[0076] Figure 6 illustrates a cache management system 600 that includes distributed lock authorities 108 within the storage appliances 106 and a local cache 1 12 stored on one of the storage appliances 106. When a storage appliance 106 first accesses a particular node, the storage appliance 106 stores that node in the local cache 1 12. As described above, when a storage appliance 106 accesses a node, the storage appliance obtains a lock from the distributed lock authority 108. When a storage appliance 106 obtains a lock from a distributed lock authority 108, the distributed lock authority 108 stores an indication (a lock entry 602) that the lock exists for a particular node and also stores a value referred to as a cache epoch 604 for that node. Each time the distributed lock authority 108 grants a write-lock for a particular node, the distributed lock authority 108 increments the cache epoch for the node by one.
[0077] When accessing a node, the distributed metadata manager 1 10
determines whether a copy stored in a local cache 1 12 is valid (if in the cache at all). To do so, the distributed metadata manager 1 10 obtains a read lock from the distributed lock authority 108 and determines the current cache epoch 604(1 ) for that node. If the cache epoch 604(1 ) read from the distributed lock authority 108 is equal to the cache epoch 604(2) stored in the local cache 1 12, then the copy of the node stored in the local cache 1 12 is valid. If the cache epochs do not match, then the copy of the node stored in the local cache is invalid. Note that even if distributed metadata manager 1 10 has a read-lock on a particular node, cache validity for that node is not guaranteed. This is because the distributed metadata manager 1 10 may have given the read-lock up at a prior time and then acquired a new read-lock on that node. In the period in which the distributed metadata manager 1 10 did not have a lock on the node, the node may have changed and thus the cached version of the node may no longer be valid.
[0078] Figure 7 illustrates an example server 700 configured to perform the functionality of the storage appliance 106, according to one embodiment of the present invention. As shown, the server 700 includes, without limitation, a central processing unit (CPU) 705, a network interface 715, a memory 720, and storage 730, each connected to a bus 717. The server 700 may also include an I/O device interface 710 connecting I/O devices 712 (e.g., keyboard, display and mouse devices) to the server 700. Further, in context of this disclosure, the computing elements shown in server 700 may correspond to a physical computing system (e.g., a system in a data center) or may be a virtual computing instance executing within a computing cloud.
[0079] The CPU 705 retrieves and executes programming instructions stored in the memory 720 as well as stores and retrieves application data residing in the storage 730. The interconnect 717 is used to transmit programming instructions and application data between the CPU 705, I/O devices interface 710, storage 730, network interface 715, and memory 720. Note that CPU 705 is included to be representative of a single CPU, multiple CPUs, a single CPU having multiple processing cores, and the like. And the memory 720 is generally included to be representative of a random access memory. The storage 730 may be a disk drive storage device. Although shown as a single unit, the storage 730 may be a combination of fixed and/or removable storage devices, such as fixed disc drives, removable memory cards, optical storage, network attached storage (NAS), or a storage area-network (SAN). Illustratively, the memory 720 includes distributed lock authority 108, a distributed metadata manager 1 10, and a distributed node manager 1 1 1 . Storage 730 includes file system metadata 1 14, local cache 1 12, and lock storage 1 16. Each element stored in memory 720 and storage 730 functions as described above.
[0080] Figure 8 is a flow diagram of method steps for creating a snapshot of a current view, according to one embodiment of the present invention. Although the method steps are described in conjunction with Figures 1 -7, persons skilled in the art will understand that any system configured to perform the method steps, in any order, falls within the scope of the present invention.
[0081 ] As shown, a method 800 begins at step 802, where distributed metadata manager 1 10 identifies the root node that is associated with the current view. A view table 220 associates the current view with a root node as described above. At step 804, distributed metadata manager 1 10 identifies the set of nodes to which the root node points. At step 806, the distributed metadata manager 1 10 creates a second root node that is a copy of the first root node and therefore points to the same set of nodes that the first root node points to, and includes the same metadata keys as the first root node. At step 808, the distributed metadata manager 1 10 generates a third root node that is a copy of the first root node and therefore points to the same set of nodes that the first root node points to, and includes the same metadata keys as the first root node. At step 810, the distributed metadata manager 1 10 associates the second root node with the current view in the view table 220. At step 812, the distributed metadata manager 1 10 associates the third root node with the snapshot in the view table 220.
[0082] Figure 9 is a flow diagram of method steps for reading, from a particular view, a value that is associated with a provided metadata key, according to one embodiment of the present invention. Although the method steps are described in conjunction with Figures 1 -7, persons skilled in the art will understand that any system configured to perform the method steps, in various alternative orders, falls within the scope of the present invention.
[0083] As shown, a method 900 begins at step 902, where the distributed metadata manager 1 10 identifies the root node that is associated with the view to be read from and acquires a read-lock on that root node. At step 903, the distributed metadata manager 1 10 identifies a node pointed to by the root node that lies on the path to the target node and sets that node as the current node. At step 904, the distributed metadata manager 1 10 acquires a read lock on the current node. At step 906, the distributed metadata manager 1 10 follows a pointer in the current node that points towards the provided metadata key and identifies the node at the end of the pointer as the "subsequent node." At step 908, the distributed metadata manager 1 10 acquires a read lock on the subsequent node.
[0084] At step 910, the distributed metadata manager 1 10 determines whether the subsequent node is the target node. The target node is the node that is the leaf node associated with the provided metadata key. If the subsequent node is the target node, then the method proceeds to step 912. At step 912, the distributed metadata manager 1 10 reads data from the subsequent node and releases all read- locks. If, at step 910, the subsequent node is not the target node, then the method proceeds to step 913. At step 913, the distributed metadata manager 1 10 determines whether there are read locks acquired for at least three nodes other than the root node. If there are read locks acquired for at least three nodes other than the root node, then the method proceeds to step 914. At step 914, the distributed metadata manager 1 10 releases the read-lock for the highest node in the tree structure. After step 914, the method proceeds to step 916. If, in step 913, the distributed metadata manager 1 10 determines that read locks are acquired for less than three nodes other than the root node, then the method proceeds to step 916. At step 916, the distributed metadata manager 1 10 sets the subsequent node to the current node. After step 916, the method returns to step 906.
[0085] Figure 10 is a flow diagram of method steps for writing, in a particular view, a value that is associated with a provided metadata key, according to one embodiment of the present invention. Although the method steps are described in conjunction with Figures 1 -7, persons skilled in the art will understand that any system configured to perform the method steps, in various alternative orders, falls within the scope of the present invention. [0086] As shown, a method 1000 begins at step 1002, where distributed metadata manager 1 10 traverses the metadata tree, acquiring read locks, towards a target node. It should be understood that for a modify-value operation, the target node is the leaf node that stores that value. For an add-key operation, the target node is the intermediate node to which distributed metadata manager 1 10 adds a leaf node as the new node. For a delete-key operation, the target node is the intermediate node whose direct child is the leaf node for deletion. While traversing, whenever the distributed metadata manager 1 10 has at least three read locks on nodes other than root nodes, the distributed metadata manager 1 10 releases the read lock that is highest up the tree.
[0087] At step 1003, the distributed metadata manager 1 10 determines whether a node is encountered that has a treelD different than that of the root node. If such a node is encountered, then the method proceeds to step 1004. If no such node is encountered, then the method proceeds to step 1012. At step 1004, the distributed metadata manager 1 10 releases the read lock on the encountered node and the parent of that node and acquires a write lock on the parent of the encountered node. At step 1006, the distributed metadata manager 1 10 determines whether the parent node is identical to before releasing the read lock. If the parent node is identical, then the method proceeds to step 1010. If the parent node is not identical, then the method proceeds to step 1008, at which the distributed metadata manager 1 10 releases all locks and restarts the write operation. After step 1008, the method returns to step 1002.
[0088] At step 1010, the distributed metadata manager acquires a read lock on the encountered node, copies the encountered node, and modifies the pointers in the parent to point to the copy of the encountered node instead of the encountered node. At step 1012, the distributed metadata manager 1 10 determines whether the encountered node (either the one that was copied if the treelDs differed or simply an encountered node with the same treelD as the root if no node was encountered with a treelD different than the root node) is the target node. If the encountered node is not the target node, then the method returns to step 1002. If the encountered node is the target node, then the method proceeds to step 1014. At step 1014, the distributed metadata manager 1 10 modifies the target node as specified in the write operation. More specifically, the distributed metadata manager 1 10 modifies, adds, or deletes a node, depending on what the write operation is.
[0089] Figure 1 1 is a flow diagram of method steps for determining whether a cache entry of a node stored in a local cache is valid, according to one embodiment of the present invention. Although the method steps are described in conjunction with Figures 1 -7, persons skilled in the art will understand that any system
configured to perform the method steps, in any order, falls within the scope of the present invention.
[0090] As shown, a method 1 100 begins at step 1 102, where the distributed metadata manager 1 10 determines whether the distributed metadata manager 1 10 has a lock acquired for the node. If the distributed metadata manager 1 10 has a lock for the node, then the method proceeds to step 1 108. If, in step 1 102, the distributed metadata manager 1 10 does not have a lock for the node, then the method proceeds to step 1 106. At step 1 106, the distributed metadata manager 1 10 acquires a read- lock on the node and reads the cache epoch for that node from the distributed lock authority 108. At step 1 108, if the cache epoch stored in the distributed lock authority 108 is equal to the cache epoch stored in the local cache, then the method proceeds to step 1 104 and determines that the cache entry is valid. If, at step 1 108, the cache epoch is not equal to the cache epoch stored in the local cache, then the method proceeds to step 108, at which the distributed metadata manager 1 10 determines that the cache entry is invalid.
[0091] One advantage of the disclosed approach is that clones of the file system metadata can be made very quickly and frequently. Another advantage is that modification operations are relatively quick as well. A further advantage is that the file system metadata is stored in a distributed manner, which allows the metadata to be stored in a robust manner.
[0092] One embodiment of the invention may be implemented as a program product for use with a computer system. The program(s) of the program product define functions of the embodiments (including the methods described herein) and can be contained on a variety of computer-readable storage media. Illustrative computer-readable storage media include, but are not limited to: (i) non-writable storage media (e.g., read-only memory devices within a computer such as CD-ROM disks readable by a CD-ROM drive, flash memory, ROM chips or any type of solid- state non-volatile semiconductor memory) on which information is permanently stored; and (ii) writable storage media (e.g., floppy disks within a diskette drive or hard-disk drive or any type of solid-state random-access semiconductor memory) on which alterable information is stored.
[0093] Embodiments of the invention are described above with reference to specific embodiments. Persons skilled in the art, however, will understand that various modifications and changes may be made thereto without departing from the broader spirit and scope of the invention as set forth in the appended claims. The foregoing description and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense.

Claims

What is claimed is:
1 . A method for performing a write operation on a set of file system
metadata, the method comprising:
receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes, and a view identifier that identifies a view;
identifying a first root node in a file system metadata tree structure that is associated with the view;
traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key; and
modifying the value in a leaf node associated with the metadata key, wherein each leaf node in the file system metadata tree store a single metadata key-value pair.
2. The method of claim 1 , wherein the first root node includes a first set of pointers, and each pointer in the first set of pointers includes an indication of which nodes are found by following that pointer.
3. The method of claim 2, wherein traversing the file system metadata tree comprises:
identifying a metadata key range that includes the metadata key and that is associated with a first pointer in the first set of pointers; identifying a first node pointed to by the first pointer; and
traversing the file system metadata tree from the first node towards the leaf node.
4. The method of claim 1 , wherein traversing the file system metadata tree comprises acquiring read locks on nodes encountered while following pointers towards the leaf node.
5. The method of claim 4, wherein traversing the file system metadata tree further comprises acquiring a first write-lock on a first parent of a first node that has a tree identifier (treelD) that is different than a treelD of the first root node and that is encountered while following the pointers towards the leaf node.
6. The method of claim 5, wherein traversing the file system metadata tree further comprises creating a copy of the first node that has the treelD that is different than the treelD of the first root node.
7. The method of claim 1 , wherein modifying the value comprises adding the leaf node to the file system metadata tree.
8. The method of claim 1 , wherein modifying the value comprises deleting the leaf node from the file system metadata tree.
9. The method of claim 1 , wherein modifying the value comprises changing the value in a metadata key-value pair stored in the leaf node.
10. The method of claim 1 , wherein the file system metadata tree includes a set of nodes that are stored in a distributed node key-value store.
1 1 . A system comprising:
a processor; and
a memory storing a metadata management application, which, when executed on the processor, performs an operation, comprising:
receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes, and a view identifier that identifies a view,
identifying a first root node in a file system metadata tree structure that is associated with the view,
traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key, and
modifying the value in a leaf node associated with the metadata key, wherein each leaf node in the file system metadata tree store a single metadata key-value pair.
12. The system of claim 1 1 , wherein the first root node includes a first set of pointers, and each pointer in the first set of pointers includes an indication of which nodes are found by following that pointer.
13. The system of claim 12, wherein traversing the file system metadata tree comprises:
identifying a metadata key range that includes the metadata key and that is associated with a first pointer in the first set of pointers; identifying a first node pointed to by the first pointer; and
traversing the file system metadata tree from the first node towards the leaf node.
14. The system of claim 1 1 , wherein traversing the file system metadata tree comprises acquiring read locks on nodes encountered while following pointers towards the leaf node.
15. The system of claim 14, wherein traversing the file system metadata tree further comprises acquiring a first write-lock on a first parent of a first node that has a tree identifier (treelD) that is different than a treelD of the first root node and that is encountered while following the pointers towards the leaf node.
16. The system of claim 15, wherein traversing the file system metadata tree further comprises creating a copy of the first node that has the treelD that is different than the treelD of the first root node.
17. The system of claim 1 1 , wherein modifying the value comprises adding the leaf node to the file system metadata tree.
18. The system of claim 1 1 , wherein modifying the value comprises deleting the leaf node from the file system metadata tree.
19. The system of claim 1 1 , wherein modifying the value comprises changing the value in a metadata key-value pair stored in the leaf node.
20. The system of claim 1 1 , wherein the file system metadata tree includes a set of nodes that are stored in a distributed node key-value store.
21 . A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the steps of:
receiving a metadata key associated with a value stored in a file system metadata tree that stores file system metadata in leaf nodes, and a view identifier that identifies a view;
identifying a first root node in a file system metadata tree structure that is associated with the view;
traversing the file system metadata tree structure from the first tree to arrive at a first node associated with the metadata key; and
modifying the value in a leaf node associated with the metadata key, wherein each leaf node in the file system metadata tree store a single metadata key- value pair.
PCT/US2015/043646 2014-08-04 2015-08-04 Write operations in a tree-based distributed file system WO2016022583A2 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US14/450,848 2014-08-04
US14/450,848 US9792309B2 (en) 2014-08-04 2014-08-04 Write operations in a tree-based distributed file system

Publications (2)

Publication Number Publication Date
WO2016022583A2 true WO2016022583A2 (en) 2016-02-11
WO2016022583A3 WO2016022583A3 (en) 2016-08-25

Family

ID=55180229

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2015/043646 WO2016022583A2 (en) 2014-08-04 2015-08-04 Write operations in a tree-based distributed file system

Country Status (2)

Country Link
US (1) US9792309B2 (en)
WO (1) WO2016022583A2 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10339016B2 (en) 2017-08-10 2019-07-02 Rubrik, Inc. Chunk allocation
US10819656B2 (en) 2017-07-24 2020-10-27 Rubrik, Inc. Throttling network bandwidth using per-node network interfaces
US11663084B2 (en) 2017-08-08 2023-05-30 Rubrik, Inc. Auto-upgrade of remote data management connectors

Families Citing this family (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9921769B2 (en) 2014-06-19 2018-03-20 Cohesity, Inc. Making more active use of a secondary storage system
EP3214551B1 (en) * 2014-11-27 2019-01-02 Huawei Technologies Co., Ltd. Metadata recovery method and apparatus
US9946604B1 (en) 2015-02-04 2018-04-17 Tintri Inc. Optimized remote cloning
US10558736B2 (en) * 2016-02-04 2020-02-11 Sap Se Metadata driven user interface layout control for web applications
US10558622B2 (en) 2016-05-10 2020-02-11 Nasuni Corporation Network accessible file server
US10169392B2 (en) 2017-03-08 2019-01-01 International Business Machines Corporation Persistent data structures on a dispersed storage network memory
US10817472B2 (en) 2017-10-23 2020-10-27 Dropbox, Inc. Storage organization system with associated storage utilization values
US10866963B2 (en) 2017-12-28 2020-12-15 Dropbox, Inc. File system authentication
CN109871715B (en) * 2019-03-12 2022-09-02 深圳市迅雷网络技术有限公司 Access method and device of distributed storage file and storage medium
CN110162525B (en) * 2019-04-17 2023-09-26 平安科技(深圳)有限公司 B+ tree-based read-write conflict resolution method, device and storage medium
CN111090824B (en) * 2019-12-23 2023-09-19 百度国际科技(深圳)有限公司 Content processing method and device
US11842060B2 (en) * 2020-09-04 2023-12-12 Cohesity, Inc. Efficiently storing data in a cloud storage
US11681454B2 (en) 2020-09-04 2023-06-20 Cohesity, Inc. Efficiently storing data in a cloud storage
CN113076292B (en) * 2021-03-30 2023-03-14 山东英信计算机技术有限公司 File caching method, system, storage medium and equipment
US11579864B1 (en) * 2021-09-20 2023-02-14 EMC IP Holding Company LLC Accelerated behavior change for upgrades in distributed systems
US11940969B2 (en) 2022-05-25 2024-03-26 International Business Machines Corporation Managing locations of derived storage objects

Family Cites Families (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7464149B2 (en) 2004-04-30 2008-12-09 International Business Machines Corporation System and method for managing introspectable objects in an enterprise
US7716182B2 (en) 2005-05-25 2010-05-11 Dassault Systemes Enovia Corp. Version-controlled cached data store
US7917474B2 (en) 2005-10-21 2011-03-29 Isilon Systems, Inc. Systems and methods for accessing and updating distributed data
US7480684B2 (en) 2005-11-04 2009-01-20 Sun Microsystems, Inc. Method and system for object allocation using fill counts
US8874627B2 (en) 2008-10-30 2014-10-28 Hewlett-Packard Development Company, L.P. Enumerating metadata in file system directories
US8572036B2 (en) 2008-12-18 2013-10-29 Datalight, Incorporated Method and apparatus for fault-tolerant memory management
US8849876B2 (en) * 2009-12-28 2014-09-30 Wenguang Wang Methods and apparatuses to optimize updates in a file system based on birth time
US8650218B2 (en) 2010-05-20 2014-02-11 International Business Machines Corporation Dynamic self configuring overlays
US9009106B1 (en) 2011-08-10 2015-04-14 Nutanix, Inc. Method and system for implementing writable snapshots in a virtualized storage environment
TWI436215B (en) 2011-10-31 2014-05-01 Delta Electronics Inc Distributed file system and method for location of replication strategy of the distributed file system
US9317511B2 (en) 2012-06-19 2016-04-19 Infinidat Ltd. System and method for managing filesystem objects
US8768977B2 (en) * 2012-07-31 2014-07-01 Hewlett-Packard Development Company, L.P. Data management using writeable snapshots in multi-versioned distributed B-trees
US8868531B2 (en) 2012-09-10 2014-10-21 Apple Inc. Concurrent access methods for tree data structures
US9037556B2 (en) 2012-12-03 2015-05-19 Vmware, Inc. Distributed, transactional key-value store
US9372757B2 (en) 2013-10-18 2016-06-21 Netapp, Inc. Incremental block level backup

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10819656B2 (en) 2017-07-24 2020-10-27 Rubrik, Inc. Throttling network bandwidth using per-node network interfaces
US11663084B2 (en) 2017-08-08 2023-05-30 Rubrik, Inc. Auto-upgrade of remote data management connectors
US10339016B2 (en) 2017-08-10 2019-07-02 Rubrik, Inc. Chunk allocation
US10423503B2 (en) 2017-08-10 2019-09-24 Rubrik, Inc. Chunk allocation
US11030062B2 (en) 2017-08-10 2021-06-08 Rubrik, Inc. Chunk allocation

Also Published As

Publication number Publication date
US9792309B2 (en) 2017-10-17
US20160034508A1 (en) 2016-02-04
WO2016022583A3 (en) 2016-08-25

Similar Documents

Publication Publication Date Title
US11892995B2 (en) Backup operations in a tree-based distributed file system
US9720926B2 (en) Read operations in a tree-based distributed file system
US9792309B2 (en) Write operations in a tree-based distributed file system
US11740808B2 (en) Making more active use of a secondary storage system
JP6798065B2 (en) Methods and equipment for efficiently implementing high-speed copyable databases
JP7090606B2 (en) Formation and operation of test data in a database system
US11714785B2 (en) Deduplicating extents across systems
US20180213035A1 (en) File replication using file content location identifiers
JP2021507341A (en) Updating the local tree for the client synchronization service
US20170315875A1 (en) Namespace policy based deduplication indexes
US9697227B2 (en) Concurrent access and transactions in a distributed file system
US9965505B2 (en) Identifying files in change logs using file content location identifiers
US10747727B2 (en) Concurrent access and transactions in a distributed file system
CN110799961B (en) System and method for creating and deleting tenants in database
US10929100B2 (en) Mitigating causality discrepancies caused by stale versioning
US10831719B2 (en) File consistency in shared storage using partial-edit files
US10732840B2 (en) Efficient space accounting mechanisms for tracking unshared pages between a snapshot volume and its parent volume
US10684992B1 (en) Causally ordering distributed file system events

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15830496

Country of ref document: EP

Kind code of ref document: A2

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 12.06.2017)

122 Ep: pct application non-entry in european phase

Ref document number: 15830496

Country of ref document: EP

Kind code of ref document: A2