PRIORITY CLAIM

[0001] This application claims priority to United States provisional patent application 62/585,054 titled "Flash Registry With On-Disk Hashing" filed on November 13, 2017, and U.S. Patent Application No. 16/121,500 titled "Flash Registry with On- Disk Hashing" filed on September 4, 2018.

BACKGROUND

[0002] Limitations and disadvantages of conventional approaches to data storage will become apparent to one of skill in the art, through comparison of such approaches with some aspects of the present method and system set forth in the remainder of this disclosure with reference to the drawings.

INCORPORATION BY REFERENCE

[0003] United States Patent Application No. 15/243,519 titled "Distributed Erasure Coded Virtual File System" is hereby incorporated herein by reference in its entirety.

BRIEF SUMMARY

[0004] Methods and systems are provided for a flash registry with on-disk hashing substantially as illustrated by and/or described in connection with at least one of the figures, as set forth more completely in the claims.

BRIEF DESCRIPTION OF THE DRAWINGS

[0005] FIG. 1 illustrates various example configurations of a virtual file system in accordance with aspects of this disclosure.

[0006] FIG. 2 illustrates an example configuration of a virtual file system node in accordance with aspects of this disclosure.

[0007] FIG. 3 illustrates another representation of a virtual file system in accordance with an example implementation of this disclosure.

[0008] FIG. 4 illustrates an example of splitting blocks of a flash registry using on- disk hashing in accordance with an example implementation of this disclosure.

[0009] FIG. 5 illustrates an example of splitting buckets of a flash registry using on- disk hashing in accordance with an example implementation of this disclosure.

[0010] FIG. 6 illustrates an example of merging blocks of a flash registry using on- disk hashing in accordance with an example implementation of this disclosure.

[0011] FIG. 7 is a flowchart illustrating a method for adding blocks of a flash registry with on-disk hashing in accordance with an example implementation of this disclosure.

[0012] FIG. 8 is a flowchart illustrating a method for accessing blocks of a flash registry with on-disk hashing in accordance with an example implementation of this disclosure.

[0013] FIG. 9 illustrates an example implementation in which two distributed failure resilient address spaces reside on a plurality of solid-state storage disks.

[0014] FIG. 10 illustrates a forward error correction scheme which may be used for protecting data stored to nonvolatile memory of a virtual file system in accordance with an example implementation of this disclosure. DETAILED DESCRIPTION

[0015] The systems in this disclosure are applicable to small clusters and can also scale to many, many thousands of nodes. An example embodiment is discussed regarding non- volatile memory (NVM), for example, flash memory that comes in the form of a solid-state drive (SSD). The NVM may be divided into 4kB "blocks" and 128MB "chunks." "Extents" may be stored in volatile memory, e.g., RAM for fast access, backed up by NVM storage as well. An extent may store pointers for blocks, e.g., 256 pointers to 1MB of data stored in blocks. In other embodiments, larger or smaller memory divisions may also be used. Metadata functionality in this disclosure may be effectively spread across many servers. For example, in cases of "hot spots" where a large load is targeted at a specific portion of the filesystem's namespace, this load can be distributed across a plurality of nodes.

[0016] FIG. 1 illustrates various example configurations of a virtual file system (VFS) in accordance with aspects of this disclosure. Shown in FIG. 1 is a local area network (LAN) 102 comprising one or more VFS nodes 120 (indexed by integers from 1 to J, for j > 1), and optionally comprising (indicated by dashed lines): one or more dedicated storage nodes 106 (indexed by integers from 1 to M, for M > 1), one or more compute nodes 104 (indexed by integers from 1 to N, for N > 1), and/or an edge router that connects the LAN 102 to a remote network 118. The remote network 118 optionally comprises one or more storage services 114 (indexed by integers from 1 to K, for K > 1), and/or one or more dedicated storage nodes 115 (indexed by integers from 1 to L, for L >

1).

[0017] Each VFS node 120 _j (j an integer, where 1 < j < J) is a networked computing device (e.g., a server, personal computer, or the like) that comprises circuitry for running VFS processes and, optionally, client processes (either directly on an operating system of the device 104 _n and/or in one or more virtual machines running in the device 104 _n ). [0018] The compute nodes 104 are networked devices that may run a VFS frontend without a VFS backend. A compute node 104 may run VFS frontend by taking an SR- IOV into the NIC and consuming a complete processor core. Alternatively, the compute node 104 may run the VFS frontend by routing the networking through a Linux kernel networking stack and using kernel process scheduling, thus not having the requirement of a full core. This is useful if a user does not want to allocate a complete core for the VFS or if the networking hardware is incompatible with the VFS requirements.

[0019] FIG. 2 illustrates an example configuration of a VFS node in accordance with aspects of this disclosure. A VFS node comprises a VFS frontend 202 and driver 208, a VFS memory controller 204, a VFS backend 206, and a VFS SSD agent 214. As used in this disclosure, a "VFS process" is a process that implements one or more of: the VFS frontend 202, the VFS memory controller 204, the VFS backend 206, and the VFS SSD agent 214. Thus, in an example implementation, resources (e.g., processing and memory resources) of the VFS node may be shared among client processes and VFS processes. The processes of the VFS may be configured to demand relatively small amounts of the resources to minimize the impact on the performance of the client applications. The VFS frontend 202, the VFS memory controller 204, and/or the VFS backend 206 and/or the VFS SSD agent 214 may run on a processor of the host 201 or on a processor of the network adaptor 218. For a multi-core processor, different VFS process may run on different cores, and may run a different subset of the services. From the perspective of the client process(es) 212, the interface with the virtual file system is independent of the particular physical machine(s) on which the VFS process(es) are running. Client processes only require driver 208 and frontend 202 to be present in order to serve them.

[0020] The VFS node may be implemented as a single tenant server (e.g., bare-metal) running directly on an operating system or as a virtual machine (VM) and/or container (e.g., a Linux container (LXC)) within a bare-metal server. The VFS may run within an LXC container as a VM environment. Thus, inside the VM, the only thing that may run is the LXC container comprising the VFS. In a classic bare-metal environment, there are user- space applications and the VFS runs in an LXC container. If the server is running other containerized applications, the VFS may run inside an LXC container that is outside the management scope of the container deployment environment (e.g. Docker).

[0021] The VFS node may be serviced by an operating system and/or a virtual machine monitor (VMM) (e.g., a hypervisor). The VMM may be used to create and run the VFS node on a host 201. Multiple cores may reside inside the single LXC container running the VFS, and the VFS may run on a single host 201 using a single Linux kernel. Therefore, a single host 201 may comprise multiple VFS frontends 202, multiple VFS memory controllers 204, multiple VFS backends 206, and/or one or more VFS drivers 208. A VFS driver 208 may run in kernel space outside the scope of the LXC container.

[0022] A single root input/output virtualization (SR-IOV) PCIe virtual function may be used to run the networking stack 210 in user space 222. SR-IOV allows the isolation of PCI Express, such that a single physical PCI Express can be shared on a virtual environment and different virtual functions may be offered to different virtual components on a single physical server machine. The I/O stack 210 enables the VFS node to bypasses the standard TCP/IP stack 220 and communicate directly with the network adapter 218. A Portable Operating System Interface for uniX (POSIX) VFS functionality may be provided through lockless queues to the VFS driver 208. SR-IOV or full PCIe physical function address may also be used to run non-volatile memory express (NVMe) driver 214 in user space 222, thus bypassing the Linux IO stack completely. NVMe may be used to access non- volatile storage media 216 attached via a PCI Express (PCIe) bus. The non-volatile storage media 216 may be, for example, flash memory that comes in the form of a solid-state drive (SSD) or Storage Class Memory (SCM) that may come in the form of an SSD or a memory module (DIMM). Other example may include storage class memory technologies such as 3D-XPoint.

[0023] The SSD may be implemented as a networked device by coupling the physical SSD 216 with the SSD agent 214 and networking 210. Alternatively, the SSD may be implemented as a network- attached NVMe SSD 242 or 244 by using a network protocol such as NVMe-oF (NVMe over Fabrics). NVMe-oF may allow access to the NVMe device using redundant network links, thereby providing a higher level or resiliency. Network adapters 226, 228, 230 and 232 may comprise hardware acceleration for connection to the NVMe SSD 242 and 244 to transform them into networked NVMe-oF devices without the use of a server. The NVMe SSDs 242 and 244 may each comprise two physical ports, and all the data may be accessed through either of these ports.

[0024] Each client process/application 212 may run directly on an operating system or may run in a virtual machine and/or container serviced by the operating system and/or hypervisor. A client process 212 may read data from storage and/or write data to storage in the course of performing its primary function. The primary function of a client process 212, however, is not storage-related (i.e., the process is only concerned that its data is reliably stored and is retrievable when needed, and not concerned with where, when, or how the data is stored). Example applications which give rise to such processes include: email servers, web servers, office productivity applications, customer relationship management (CRM), animated video rendering, genomics calculation, chip design, software builds, and enterprise resource planning (ERP).

[0025] A client application 212 may make a system call to the kernel 224 which communicates with the VFS driver 208. The VFS driver 208 puts a corresponding request on a queue of the VFS frontend 202. If several VFS frontends exist, the driver may load balance accesses to the different frontends, making sure a single file/directory is always accessed via the same frontend. This may be done by "sharding" the frontend based on the ID of the file or directory. The VFS frontend 202 provides an interface for routing file system requests to an appropriate VFS backend based on the bucket that is responsible for that operation. The appropriate VFS backend may be on the same host or it may be on another host.

[0026] The VFS backend 206 hosts several buckets, each one of them services the file system requests that it receives and carries out tasks to otherwise manage the virtual file system (e.g., load balancing, journaling, maintaining metadata, caching, moving of data between tiers, removing stale data, correcting corrupted data, etc.)

[0027] The VFS SSD agent 214 handles interactions with a respective storage device 216. This may include, for example, translating addresses, and generating the commands that are issued to the storage device (e.g., on a SATA, SAS, PCIe, or other suitable bus). Thus, the VFS SSD agent 214 operates as an intermediary between a storage device 216 and the VFS backend 206 of the virtual file system. The SSD agent 214 could also communicate with a standard network storage device supporting a standard protocol such as NVMe-oF (NVMe over Fabrics).

[0028] FIG. 3 illustrates another representation of a virtual file system in accordance with an example implementation of this disclosure. In FIG. 3, the element 302 represents memory resources (e.g., DRAM and/or other short-term memory) and processing (e.g., x86 processor(s), ARM processor(s), NICs, ASICs, FPGAs, and/or the like) resources of various node(s) (compute, storage, and/or VFS) on which resides a virtual file system, such as described regarding FIG. 2 above. The element 308 represents the one or more physical storage devices 216 which provide the long term storage of the virtual file system.

[0029] As shown in FIG. 3, the physical storage is organized into a plurality of distributed failure resilient address spaces (DFRASs) 518. Each of which comprises a plurality of chunks 310, which in turn comprises a plurality of blocks 312. The organization of blocks 312 into chunks 310 is only a convenience in some implementations and may not be done in all implementations. Each block 312 stores committed data 316 (which may take on various states, discussed below) and/or metadata 314 that describes or references committed data 316.

[0030] The organization of the storage 308 into a plurality of DFRASs enables high performance parallel commits from many - perhaps all - of the nodes of the virtual file system (e.g., all nodes 104 ₁ -104 _N , ΙΟό _ϊ -ΙΟό _Μ , and 120^120; of FIG. 1 may perform concurrent commits in parallel). In an example implementation, each of the nodes of the virtual file system may own a respective one or more of the plurality of DFRAS and have exclusive read/commit access to the DFRASs that it owns.

[0031] Each bucket owns a DFRAS, and thus does not need to coordinate with any other node when writing to it. Each bucket may build stripes across many different chunks on many different SSDs, thus each bucket with its DFRAS can choose what "chunk stripe" to write to currently based on many parameters, and there is no coordination required in order to do so once the chunks are allocated to that bucket. All buckets can effectively write to all SSDs without any need to coordinate.

[0032] Each DFRAS being owned and accessible by only its owner bucket that runs on a specific node allows each of the nodes of the VFS to control a portion of the storage 308 without having to coordinate with any other nodes (except during [re] assignment of the buckets holding the DFRASs during initialization or after a node failure, for example, which may be performed asynchronously to actual reads/commits to storage 308). Thus, in such an implementation, each node may read/commit to its buckets' DFRASs independently of what the other nodes are doing, with no requirement to reach any consensus when reading and committing to storage 308. Furthermore, in the event of a failure of a particular node, the fact the particular node owns a plurality of buckets permits more intelligent and efficient redistribution of its workload to other nodes (rather the whole workload having to be assigned to a single node, which may create a "hot spot")- In this regard, in some implementations the number of buckets may be large relative to the number of nodes in the system such that any one bucket may be a relatively small load to place on another node. This permits fine grained redistribution of the load of a failed node according to the capabilities and capacity of the other nodes (e.g., nodes with more capabilities and capacity may be given a higher percentage of the failed nodes buckets).

[0033] To permit such operation, metadata may be maintained that maps each bucket to its current owning node such that reads and commits to storage 308 can be redirected to the appropriate node.

[0034] Load distribution is possible because the entire filesystem metadata space (e.g., directory, file attributes, content range in the file, etc.) can be broken (e.g., chopped or sharded) into small, uniform pieces (e.g., "shards")- For example, a large system with 30k servers could chop the metadata space into 128k or 256k shards.

[0035] Each such metadata shard may be maintained in a "bucket." Each VFS node may have responsibility over several buckets. When a bucket is serving metadata shards on a given backend, the bucket is considered "active" or the "leader" of that bucket. Typically, there are many more buckets than VFS nodes. For example, a small system with 6 nodes could have 120 buckets, and a larger system with 1,000 nodes could have 8k buckets.

[0036] Each bucket may be active on a small set of nodes, typically 5 nodes that that form a penta-group for that bucket. The cluster configuration keeps all participating nodes up-to-date regarding the penta-group assignment for each bucket.

[0037] Each penta-group monitors itself. For example, if the cluster has 10k servers, and each server has 6 buckets, each server will only need to talk with 30 different servers to maintain the status of its buckets (6 buckets will have 6 penta-groups, so 6*5 = 30). This is a much smaller number than if a centralized entity had to monitor all nodes and keep a cluster-wide state. The use of penta-groups allows performance to scale with bigger clusters, as nodes do not perform more work when the cluster size increases. This could pose a disadvantage that in a "dumb" mode a small cluster could actually generate more communication than there are physical nodes, but this disadvantage is overcome by sending just a single heartbeat between two servers with all the buckets they share (as the cluster grows this will change to just one bucket, but if you have a small 5 server cluster then it will just include all the buckets in all messages and each server will just talk with the other 4). The penta-groups may decide (i.e., reach consensus) using an algorithm that resembles the Raft consensus algorithm.

[0038] Each bucket may have a group of compute nodes that can run it. For example, five VFS nodes can run one bucket. However, only one of the nodes in the group is the controller/leader at any given moment. Further, no two buckets share the same group, for large enough clusters. If there are only 5 or 6 nodes in the cluster, most buckets may share backends. In a reasonably large cluster there many distinct node groups. For

26'

example, with 26 nodes, there are more than 64,000 ( _S ^ ₂₆ _ _s y ) possible five-node groups (i.e., penta-groups).

[0039] All nodes in a group know and agree (i.e., reach consensus) on which node is the actual active controller (i.e., leader) of that bucket. A node accessing the bucket may remember ("cache") the last node that was the leader for that bucket out of the (e.g., five) members of a group. If it accesses the bucket leader, the bucket leader performs the requested operation. If it accesses a node that is not the current leader, that node indicates the leader to "redirect" the access. If there is a timeout accessing the cached leader node, the contacting node may try a different node of the same penta-group. All the nodes in the cluster share common "configuration" of the cluster, which allows the nodes to know which server may run each bucket. [0040] Each bucket may have a load/usage value that indicates how heavily the bucket is being used by applications running on the filesystem. For example, a server node with 11 lightly used buckets may receive another bucket of metadata to run before a server with 9 heavily used buckets, even though there will be an imbalance in the number of buckets used. Load value may be determined according to average response latencies, number of concurrently run operations, memory consumed or other metrics.

[0041] Redistribution may also occur even when a VFS node does not fail. If the system identifies that one node is busier than the others based on the tracked load metrics, the system can move (i.e., "fail over") one of its buckets to another server that is less busy. However, before actually relocating a bucket to a different host, load balancing may be achieved by diverting writes and reads. Since each write may end up on a different group of nodes, decided by the DFRAS, a node with a higher load may not be selected to be in a stripe to which data is being written. The system may also opt to not serve reads from a highly loaded node. For example, a "degraded mode read" may be performed, wherein a block in the highly loaded node is reconstructed from the other blocks of the same stripe. A degraded mode read is a read that is performed via the rest of the nodes in the same stripe, and the data is reconstructed via the failure protection. A degraded mode read may be performed when the read latency is too high, as the initiator of the read may assume that that node is down. If the load is high enough to create higher read latencies, the cluster may revert to reading that data from the other nodes and reconstructing the needed data using the degraded mode read.

[0042] Each bucket manages its own distributed erasure coding instance (i.e., DFRAS 518) and does not need to cooperate with other buckets to perform read or write operations. There are potentially thousands of concurrent, distributed erasure coding instances working concurrently, each for the different bucket. This is an integral part of scaling performance, as it effectively allows any large filesystem to be divided into independent pieces that do not need to be coordinated, thus providing high performance regardless of the scale.

[0043] Each bucket handles all the file systems operations that fall into its shard. For example, the directory structure, file attributes and file data ranges will fall into a particular bucket's jurisdiction.

[0044] An operation done from any frontend starts by finding out what bucket owns that operation. Then the backend leader, and the node, for that bucket is determined. This determination may be performed by trying the last-known leader. If the last-known leader is not the current leader, that node may know which node is the current leader. If the last-known leader is not part of the bucket's penta-group anymore, that backend will let the front end know that it should go back to the configuration to find a member of the bucket's penta-group. The distribution of operations allows complex operations to be handled by a plurality of servers, rather than by a single computer in a standard system.

[0045] An on-disk hash (ODH) may manage hash tables implemented over 4k registry pages. These hash tables may have keys and values (e.g., <KEY, VALUE> pairs, key- value entries) of various sizes, supporting many other higher level data- structures that run inside (and across) buckets.

[0046] The disclosed system may support many hash tables of the same kind (with a different object ID) stored over the same registry or different kind of hash tables stored over the same registry. The hash may be based on a cryptographic hashing function and provides a unique signature of each hash kind and instance. The registry key for the ODH may be based on the type of hash, the ID of hash instance, and a block index. Because the registry (comprising e.g., block indices) is stored in RAM, it is efficient to look for a 4k block and find out whether it exists or not. [0047] The ODH is an expandable (and collapsible) hash implemented efficiently over 4k blocks that are stored in the registry. Therefore, the number of 4k blocks used is proportional to the amount data that is stored on the ODH.

[0048] The ODH has dynamic structures based on the KEY and VALUE types, and stores the hash information as a sorted <KEY, VALUE> array that allows for an efficient binary search over. The sorting is done based on a cryptographic hash signature of the key and not the key itself, to ensure uniform spread of the sorting criterion across all bits. Because the hash set may consume more than what would fit in a single 4k block, a mechanism for expanding and collapsing the hash set is required.

[0049] FIG. 4 illustrates an example of splitting blocks of a flash registry using ODH in accordance with an example implementation of this disclosure. A table of 400 entries and 10 bytes per entry (e.g., two bytes for key, 8 bytes for value) may be implemented in each 4kB ODH registry block. For illustration purposes, the registry blocks in FIG. 4 is shown with a maximum of four <KEY, VALUE> entries. In addition to the local hash table of <KEY, VALUE> pairs, each ODH registry block may comprise header information that indicates an index and a split level. The split level may be stored inside the block apart from the "key" (or hash of the block).

[0050] The elements of each ODH block are sorted by the key and may be serialized one after the other.

[0051] The ODH registry blocks are stored over the DFRAS. Set/get operations are implemented by the ODH, which is several layers "higher" than the DFRAS. The ODH assumes that there is a functioning registry that provides the service of storing the ODH 4k data based on the index for each such page. This also achieves write leveling when an ODH page is updated. [0052] As long as the number of entries fits within one block, the code may look up the ODH by searching a single block. With the type and instance ID, the ODH may enumerate the <KEY, VALUE> pairs in the block until it finds the right pair.

[0053] When the block fills up, it increases the split level, and moves some of the entries to a new block based on the cryptographic hash. The entries are moved to the blocks by the least significant bits (LSBs) of the hash of the key. The LSBs that match the index remain, and the LSB that match the new block (with added "1") move to the new block. Because sorting and moving is based on the hash of the keys and such hashes generate a uniform distribution of hashed values, about half of the entries would remain in the original block, and half of the entries would be migrated to the newly created block.

[0054] The ODH may be created using one registry block 401 A without a hash index at split level 0 <index=NIL, split=l>. The LSBs of the keys in the first four ODH entries are 0010, 1010, 1110, and 1111, and the corresponding values are value_x0, value_xl, value_x2, and value_x3 respectively.

[0055] When splitting blocks, a new block is added and the existing block does not change its address. This maintains an "anchor" for finding data. After a split, the anchor block is the block with an index that begins with "0", because when adding a leading "0" to a number the value of that number does not change, thus the address of that block does not actually change.

[0056] In FIG. 4, when registry block 401 A is split, a new block 402A is added with header information <index=lb, split=l>. The <KEY, VALUE> pairs in existing block 401 A that end in "1" are moved into the new block 402A. The <KEY, VALUE> pairs in existing block 401 A that end in "0" remain in the hash table, and the header information is changed from <index=NIL, split=0> to <index=0b, split=l>. For illustration purposes, after block 401 A is split, it is designated as block 401B, although the physical memory address of the block does not change.

[0057] After block 401 A is split, ODH registry block 40 IB receives <KEY, VALUE> pair (0000, value_y0). Because registry block 40 IB has reached its capacity, registry block 40 IB is split and a new block 403 A is added with header information <index=10b, split=2>. The <KEY, VALUE> pairs in existing block 401B that end in "10" are moved into the new block 403 A. The <KEY, VALUE> pairs in existing block 401 A that end in "00" remain in the hash table, and the header information is changed from <index=0b, split=l> to <index=00b, split=2>. For illustration purposes, after block 40 IB is split, it is designated as block 401C.

[0058] After being split from block 401A, ODH registry block 402A receives <KEY, VALUE> pairs (0001, value_yl), (0101, value_y2) and (1011, value_y3). Because registry block 402A has reached its capacity, registry block 402A is split and a new block 404A is added with header information <index=l lb, split=2>. The <KEY, VALUE> pairs in existing block 402A that end in "11" are moved into the new block 404A. The <KEY, VALUE> pairs in existing block 401 A that end in "01" remain in the hash table, and the header information is changed from <index=lb, split=l> to <index=01b, split=2>. For illustration purposes, after block 402A is split, it is designated as block 402B.

[0059] After being split from block 401B, ODH registry block 403 A receives <KEY, VALUE> pair (0110, value_z3). Because registry block 403A has reached its capacity, registry block 403A is split and a new block 405A is added. The <KEY, VALUE> pairs in existing block 403a that end in "110" are moved into the new block 405A with header information <index=110b, split=3>. The <KEY, VALUE> pairs in existing block 403A that end in "010" remain in the hash table, and the header information is changed from <index=10b, split=2> to <index=010b, split=3>. For illustration purposes, after block 403 A is split, it is designated as block 403B.

[0060] After block 40 IB is split, ODH registry block 401C receives <KEY, VALUE> pairs (0100, value_z0) and (1000, value_zl). Because registry block 401C has not reached its capacity, splitting is not required.

[0061] FIG. 4 illustrates an ODH in a single bucket that is handled by a single backend (and a single CPU core). FIG. 5 illustrates an example of splitting buckets of a flash registry using on-disk hashing in accordance with an example implementation of this disclosure. To support hash sets that are very big, hash set handling may be supported over several processes (and several buckets). Further, to implement load balancing, the buckets themselves may be split and or merged. For example, when new servers are added to a cluster, existing buckets may be split and redistributed across all of the servers.

[0062] Following the splits described regarding FIG. 4, the ODH comprises block 401C, block 403B, block 405 A, block 402B, and block 404A. All five of these ODH blocks may be located on a single bucket 500A. For illustration purposes, bucket 500A and bucket 500B are the same logical buckets. Bucket 500B represents bucket 500A after a bucket split.

[0063] When a bucket split condition has been determined for the ODH, the ODH of bucket 500 A can be split into buckets 500B and 501 A, where bucket 501 A is a new bucket.

[0064] A strong property of this bucket splitting is that just half of the 4k blocks have to move to a new bucket, and half of the blocks remain. This is true for both splitting the ODH and splitting the whole bucket. If the ODH is split as in FIG. 4, half blocks relocate to another existing bucket. If a bucket with all its ODH's are split, each ODH will leave half of the <KEY, VALUE> entries with the original bucket, and half of the <KEY, VALUE> entries will be relocated to the new bucket.

[0065] The split may be determined according to the LSBs of the keys in each block. For illustration purposes, after a bucket is split, the corresponding blocks are designated with a new version number (e.g., 401C becomes 401D), although the physical memory address of the blocks may not change. Additionally, the split level (corresponding to the index width) is reduced.

[0066] The LSB of the keys in blocks 401C, 403B and 405A is "0" (corresponding to bucket 500). These blocks are relabeled as blocks 401D, 403C and 405B respectively and remain in the same bucket (i.e., bucket 500A relabeled as bucket 500B) after the ODH on bucket 500A is split.

[0067] The LSB of the keys in blocks 402B and 404A is "1" (corresponding new bucket 501). These blocks are relabeled as blocks 402C and 404B respectively and are moved to bucket 501 A (bucket "1") after bucket 500A is split.

[0068] When the ODH that now resides on buckets 500B and 501 A is split further, two LSBs of the keys are considered. The two LSBs of the keys in block 40 ID are "00" (corresponding to bucket 500). Block 40 ID is relabeled as block 40 IE and remains in the same bucket (i.e., bucket 500B relabeled as bucket 500C) after the ODH on bucket 500B (bucket "0") is split.

[0069] The two LSBs of the keys in blocks 403C and 405B are "10" (corresponding to new bucket 502). Blocks 403C and 405B are relabeled as blocks 403D and 405C and are moved to bucket 502A (bucket "10") after the ODH on bucket 500B (bucket "0") is split.

[0070] The two LSBs of the keys in block 402C are "01" (corresponding to bucket 501). Block 402C is relabeled as block 402D and remains in the same bucket (i.e., bucket 501A relabeled as bucket 501B) after the ODH on bucket 501A (bucket "1") is split.

[0071] The two LSBs of the keys in block 404B are "11" (corresponding to new bucket 503). Block 404B is relabeled as block 404C and is moved to bucket 503A (bucket "11") after the ODH on bucket 501 A (bucket "1") is split.

[0072] If the combined number of entries would fit in a small enough portion of a block (e.g., less than 2k of a 4k block), sibling blocks (i.e., two blocks that are at the same split level and are from the same parent block) may be merged. FIG. 6 illustrates an example of merging blocks of a flash registry using on-disk hashing in accordance with an example implementation of this disclosure.

[0073] Following the splits described regarding FIG. 4, the ODH comprises block 401C, block 403B, block 405A, block 402B, and block 404A. Block 403B and block 405A are at the same split level (i.e., split level 3) and are from the same parent block. If, for example, two <KEY, VALUE> entries are removed from block 405 A, the combined number of entries (i.e., 2) would fit in a small enough portion of a block. Therefore, the split level and corresponding index width of block 403B may be reduced to form block 601 (which may be at the same address as 403B).

[0074] Likewise, block 402B and block 404A are at the same split level (i.e., split level 2) and are from the same parent block. If, for example, a <KEY, VALUE> entry is removed from block 402B and two <KEY, VALUE> entries are removed from block 404A, the combined number of entries (i.e., 2) would fit in a small enough portion of a block. Therefore, the split level and corresponding index width of block 402B may be reduced to form block 602 (which may be at the same address as 402B).

[0075] FIG. 7 is a flowchart illustrating an example method for adding blocks of a flash registry with on-disk hashing in accordance with an example implementation of this disclosure. At block 701, a <KEY, VALUE> pair to be added to registry is received by the ODH.

[0076] At block 703, the corresponding bucket and block are located according to one or more LSBs of the key. The number of LSBs may be determined according to the current largest split level. If the one or more LSBs do not match an index, the number of LSBs may be reduced by one. The search for a matching index is continued over the blocks having the second largest split level, and so on. For example, if index = "11" (assuming the largest split level is 2) is not found, the search continues for index = "1" at split level = 1.

[0077] If adding an additional <KEY, VALUE> entry to the appropriate table, exceeds or nears block capacity at 705, a block split may be necessary. If a new block can be added at 707, a block split may occur at 711. If a new block cannot be added at 707, an IO error may be returned at 709. Reasons why a new block cannot be added include memory capacity and/or processor load.

[0078] If the block capacity at 705 was determined to be sufficient or a block split was made at 711, the newly received <KEY, VALUE> pair is added at 713. At 715, a bucket split, as described in FIG. 5, can be made if the DEC indicates that it is necessary.

[0079] FIG. 8 is a flowchart illustrating a method for accessing blocks of a flash registry with ODH in accordance with an example implementation of this disclosure. At block 801, a key hash is calculated.

[0080] At block 803, the corresponding bucket and block are searched for according to one or more LSBs of the key. The number of LSBs may be determined according to the current largest split level. If the one or more LSBs do not match an index, the number of LSBs may be reduced by one. The search for a matching index is continued over the blocks having the second largest split level, and so on. For example, if index = "11" (assuming the largest split level is 2) is not found, the search continues for index = "1" at split level = 1. If an entry corresponding to the key is not found at 803, an IO error may be returned at 805.

[0081] If an entry corresponding to the key is found at 803, the <KEY, VALUE> entry in the hash table is accessed at 807. If accessed <KEY, VALUE> entry is removed at 809, there is a possibility for merging blocks. If the total number of <KEY, VALUE> entries in the accessed table and its sibling table fit within a single block at 811, the table and its sibling table are merged into the anchor block at 813.

[0082] FIG. 9 illustrates an example implementation in which two distributed failure resilient address spaces reside on a plurality of solid-state storage disks. The chunks 510 _1;1 -510 _D; c may be organized into a plurality of chunk stripes 520^520 ₃ (S being an integer). In an example implementation, each chunk stripe 520 _s (s being an integer, where 1 < s < S) is separately protected using forward error correction (e.g., erasure coding). The number of chunks 510 _{d c} in any particular chunk stripe 520 _s may thus be determined based on the desired level of data protection.

[0083] Assuming, for purposes of illustration, that each chunk stripe 520 _s comprises N = M + K (where each of N, M, and K are integers) chunks 510 _d,c , then M of the N chunks 510 _{d c} may store data digits (typically binary digits or "bits" for current storage devices) and K of the N chunks 510 _d,c may store protection digits (again, typically bits). To each stripe 520 _s , then, the virtual file system may assign N chunks 510 _d,c from N different failure domains.

[0084] As used herein, a "failure domain" refers to a group of components in which a failure of any single one of the components (the component losing power, becoming nonresponsive, and/or the like) may result in failure of all the components. For example, if a rack has a single top-of-the-rack switch a failure of that switch will bring down connectivity to all the components (e.g., compute, storage, and/or VFS nodes) on that rack. Thus, to the rest of the system it is equivalent to if all of the components on that rack failed together. A virtual file system in accordance with this disclosure may comprise fewer failure domains than chunks 510.

[0085] In an example implementation where the nodes of the virtual file system are connected and powered in a fully-redundant way with only a single storage device 506 per such node, a failure domain may be just that single storage device 506. Thus, in an example implementation, each chunk stripe 520 _s comprises a plurality of chunks 510 _{d c} residing on each of N of the storage devices 506^506 ₀ -, (D is thus greater than or equal to N). An example of such an implementation is shown in FIG 7.

[0086] In FIG. 9, D = 7, N = 5, M = 4, K = 1, and the storage is organized into two DFRASs. These numbers are merely for illustration and not intended as limiting. Three chunk stripes 520 of the first DFRAS are arbitrarily called out for illustration. The first chunk stripe 520i consists of chunks 510 _1;1 , 510 _2;2 , 510 _3;3 , 510 _4;5 and 510 _5;6 ; the second chunk stripe 520 ₂ consists of chunks 510 _{3 2} , 510 _4;3 , 510 _{5 3} , 510 ₆₂ and 510 _7;3 ; and the third chunk stripe 520 ₃ consists of chunks 510 _1;4 , 510 _2;4 , 510 _3;5 , 510 _5;7 and 510 _7;5 .

[0087] Although D = 7 and N = 5 in the simple example of FIG. 9 in an actual implementation D may be much larger than N (e.g., by a multiple of an integer greater than 1 and possibly as high as many orders of magnitude) and the two values may be chosen such that the probability of any two chunk stripes 520 of a single DFRAS residing on the same set of N storage devices 506 (or, more generally, on the same set of N failure domains) is below a desired threshold. In this manner, failure of any single storage device 506 _d (or, more generally, any single failure domain) will result (with the desired statistical probably determined based on: chosen values of D and N, the sizes of the N storage devices 506, and the arrangement of failure domains) in loss of at most one chunk 510 _{b c} of any particular stripe 520 _s . Even further, a dual failure will result in vast majority of stripes losing at most a single chunk 510 _b,c and only small number of stripes (determined based on the values of D and N) will lose two chunks out of any particular stripe 520 _s (e.g., the number of two-failure stripes may be exponentially less than the number of one-failure stripes).

[0088] For example, if each storage device 506 _d is 1TB, and each chunk is 128MB, then failure of storage device 506 _d will result (with the desired statistical probably determined based on: chosen values of D and N, the sizes of the N storage devices 506, and the arrangement of failure domains) in 7812 (= 1TB/128MB) chunk stripes 520 losing one chunk 510. For each such affected chunk stripe 520 _s , the lost chunk 510 _d,c can be quickly reconstructed using an appropriate forward error correction algorithm and the other N-l chunks of the chunk stripe 520 _s . Furthermore, since the affected 7812 chunk stripes are uniformly distributed across all of the storage devices 506^506 ₀ , reconstructing the lost 7812 blocks 510 _d,c will involve (with the desired statistical probably determined based on: chosen values of D and N, the sizes of the N storage devices 506, and the arrangement of failure domains) reading the same amount of data from each of storage devices 506i-506 _D (i.e., the burden of reconstructing the lost data is uniformly distributed across all of storage devices 506^506 ₀ so as to provide for very quick recovery from the failure).

[0089] Next, turning to the case of a concurrent failure of two of the storage devices 506i-506 _D (or, more generally, concurrent failure of two failure domains), due to the uniform distribution of the chunk stripes 520i-520s of each DFRAS over all of the storage devices 506^506 ₀ , only a very small number of chunk stripes 520^520 ₃ will have lost two of their N chunks. The virtual file system may be operable to quickly identify such two-loss chunk stripes based on metadata which indicates a mapping between chunk stripes 520i-520s and the storage devices 506^506 ₀ - Once such two- loss chunk stripes are identified, the virtual file system may prioritize reconstructing those two-loss chunk stripes before beginning reconstruction of the one-loss chunk stripes. The remaining chunk stripes will have only a single lost chunk and for them (the vast majority of the affected chunk stripes) a concurrent failure of two storage devices 506 _d is the same as a failure of only one storage device 506 _d . Similar principles apply for a third concurrent failure (the number of chunk stripes having three failed blocks will be even less than the number having two failed blocks in the two concurrent failure scenario), and so on. In an example implementation, the rate at which reconstruction of a chunk stripe 520 _s is performed may be controlled based on the number of losses in the chunk stripe 520 _s . This may be achieved by, for example, controlling the rates at which reads and commits for reconstruction are performed, the rates at which FEC computations for reconstruction are performed, the rates at which network messages for reconstruction are communicated, etc.

[0090] FIG. 10 illustrates a forward error correction scheme which may be used for protecting data stored to nonvolatile memory of a virtual file system in accordance with an example implementation of this disclosure. Shown are storage blocks 902 _1;1 -902 _7;7 of block stripes 530i-530 ₄ of a DFRAS. In the protection scheme of FIG. 10, five blocks of each stripe are for storage of data digits and two blocks of each stripe are for data storage of protection digits (i.e., M=5 and K=2). In FIG. 10, the protection digits are calculated using the following equations (l)-(9):

PI = D1 ₁ 0D22®D33®D44®D54 (1)

P2 = D2 ₁ eD3 ₂ eD43eD5 ₃ eDl4 (2)

P3 = D3 ₁ eD4 ₂ eD5 ₂ eDl3eD2 ₄ (3)

P4 = D4 ₁ eD5 ₁ eDl ₂ eD2 ₃ eD3 ₄ (4)

Qi = Di ₁ eDi ₂ eDi ₃ eDi ₄ ez (6)

Q2 = D2 _! eD2 ₂ eD2 ₃ eD2 ₄ eZ (7) Q3 = D3 _! eD32eD33eD3 ₄ eZ (8) Q4 = D4 _! eD4 ₂ eD43eD4 ₄ eZ (9)

[0091] Thus, the four stripes 530i-530 ₄ in FIG. 10 are part of a multi-stripe (four stripes, in this case) FEC protection domain and loss of any two or fewer blocks in any of the block stripes 530i-530 ₄ can be recovered from using various combinations of the above equations (l)-(9). For comparison, an example of a single-stripe protection domain would be if Dl _b D22, D33, D4 ₄ , D5 ₄ were protected only by PI and Dl _b D22, D33, D4 ₄ , D5 ₄ , and PI were all written to stripe 530i (530i would be the single-stripe FEC protection domain).

[0092] In accordance with an example implementation of this disclosure, a plurality of computing devices are communicatively coupled to each other via a network, and each of the plurality of computing devices comprises one or more of a plurality of storage devices. A plurality of failure resilient address spaces are distributed across the plurality of storage devices such that each of the plurality of failure resilient address spaces spans a plurality of the storage devices. Each one of the plurality of failure resilient address spaces is organized into a plurality of stripes (e.g., a plurality of 530 as in FIGs. 9 and 10). Each one or more stripes of the plurality of stripes is part of a respective one of a plurality of forward error correction (FEC) protection domains (e.g., a multi-stripe FEC domain such as in FIG. 10). Each of the plurality of stripes may comprise a plurality of storage blocks (e.g., a plurality of 512). Each block of a particular one of the plurality of stripes may reside on a different one of the plurality of storage devices. A first portion the plurality of storage blocks (e.g., the quantity of five consisting of 902 _1;2 -902 _1;6 of stripe 530i of FIG. 10) may be for storage of data digits, and a second portion of the plurality of storage blocks (e.g., the quantity of two 902 and 902 _{1 7} of stripe 530i of FIG. 10) may be for storage of protection digits calculated based, at least in part, on the data digits. The plurality of computing devices may be operable to rank the plurality of stripes. The rank may be used for selection of which of the plurality of stripes to use for a next commit operation to the one of the plurality of failure resilient address spaces. The rank may be based on how many protected and/or unprotected storage blocks are in each of the plurality of stripes. For any particular one of the plurality of stripes, the rank may be based on a bitmap stored on the plurality of storage devices with the particular one of the plurality of stripes. The rank may be based on how many blocks currently storing data are in each of the plurality of stripes. The rank may be based on read and write overhead for committing to each of the plurality of stripes. Each of the failure resilient address spaces may be owned by only one of the plurality of computing devices at any given time, and each one of the plurality of failure resilient address spaces may be read and written only by its owner. Each of the computing devices may own multiple of the failure resilient address spaces. The plurality of storage devices may be organized into a plurality of failure domains. Each one of the plurality of stripes may span a plurality of the failure domains. Each of the failure resilient address spaces may span all of the plurality of failure domains, such that upon failure of any particular one of the plurality of failure domains, a workload for reconstructing the lost data is distributed among each of the others of the plurality of failure domains. The plurality of stripes may be distributed across the plurality of failure domains such that, in the event of concurrent failure of two of the plurality of failure domains, the chances of two blocks of any particular one of the plurality stripes residing on the failed two of the plurality of failure domains is exponentially less than the chances of only one block of any particular one of the plurality stripes residing on the failed two of the plurality of failure domains. The plurality of computing devices may be operable to first reconstruct any of the plurality of stripes which have two failed blocks, and then reconstruct any of the plurality of stripes which have only one failed block. The plurality of computing devices may be operable to perform the reconstruction of the plurality of stripes which have two failed blocks at a higher rate (e.g., with a greater percentage of CPU clock cycles dedicated to the reconstruction, a greater percentage of network transmit opportunities dedicated to the reconstruction, and/or the like.) than the rate of reconstruction of the plurality of stripes which have only one failed block. The plurality of computing devices may be operable to determine, in the event of a failure of one or more of the failure domains, a rate at which any particular lost block is reconstructed based on how many other blocks of a same one of the plurality of stripes have been lost. Wherein one or more of the plurality of failure domains comprises a plurality of the storage devices. Each of the plurality of FEC protection domains may span multiple stripes of the plurality of stripes. The plurality of stripes may be organized into a plurality of groups (e.g., chunk stripes 52(^-520 ₃ as in FIG. 9), where each of the plurality of groups comprises one or more of the plurality of stripes, and, the plurality of computing devices are operable to rank, for each of the groups, the one or more of the plurality of stripes of the group. The plurality of computing devices may be operable to: perform successive committing operations to a selected one of the plurality of groups until the one or more of the plurality of stripes of the of the group no longer meets a determined criterion, and upon the selected one of the plurality of groups no longer meeting the determined criterion, select a different one of the plurality of groups. The criterion may be based on how many blocks are available for new data to be written to.

[0093] While the present method and/or system has been described with reference to certain implementations, it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted without departing from the scope of the present method and/or system. In addition, many modifications may be made to adapt a particular situation or material to the teachings of the present disclosure without departing from its scope. Therefore, it is intended that the present method and/or system not be limited to the particular implementations disclosed, but that the present method and/or system will include all implementations falling within the scope of the appended claims. [0094] As utilized herein the terms "circuits" and "circuitry" refer to physical electronic components (i.e. hardware) and any software and/or firmware ("code") which may configure the hardware, be executed by the hardware, and or otherwise be associated with the hardware. As used herein, for example, a particular processor and memory may comprise first "circuitry" when executing a first one or more lines of code and may comprise second "circuitry" when executing a second one or more lines of code. As utilized herein, "and/or" means any one or more of the items in the list joined by "and/or". As an example, "x and/or y" means any element of the three-element set {(x), (y), (x, y)}. In other words, "x and/or y" means "one or both of x and y". As another example, "x, y, and/or z" means any element of the seven-element set {(x), (y), (z), (x, y), (x, z), (y, z), (x, y, z)}. In other words, "x, y and/or z" means "one or more of x, y and z". As utilized herein, the term "exemplary" means serving as a non-limiting example, instance, or illustration. As utilized herein, the terms "e.g.," and "for example" set off lists of one or more non-limiting examples, instances, or illustrations. As utilized herein, circuitry is "operable" to perform a function whenever the circuitry comprises the necessary hardware and code (if any is necessary) to perform the function, regardless of whether performance of the function is disabled or not enabled (e.g., by a user- configurable setting, factory trim, etc.).