Skip to content

Latest commit

 

History

History
187 lines (137 loc) · 6.74 KB

ENGINES-experimental.md

File metadata and controls

187 lines (137 loc) · 6.74 KB
Raw

Experimental Storage Engines for pmemkv

tree3

A persistent single-threaded engine, backed by a read-optimized B+ tree. It is disabled by default. It can be enabled in CMake using the ENGINE_TREE3 option.

Configuration

Configuration must specify a path to a PMDK persistent pool, which can be a file (on a DAX filesystem), a DAX device, or a PMDK poolset file.

  • path -- Path to the database file
    • type: string
  • force_create -- If 0, pmemkv opens file specified by 'path', otherwise it creates it
    • type: uint64_t
    • default value: 0
  • size -- Only needed when force_create is not 0, specifies size of the database [in bytes]
    • type: uint64_t
    • min value: 8388608 (8MB)

Internals

Internally, tree3 uses a hybrid fingerprinted B+ tree implementation. Rather than keeping inner and leaf nodes of the tree in persistent memory, tree3 uses a hybrid structure where inner nodes are kept in DRAM and leaf nodes only are kept in persistent memory. Though tree3 has to recover all inner nodes when the engine is started, searches are performed in DRAM except for a final read from persistent memory.

pmemkv-intro

Leaf nodes in tree3 contain multiple key-value pairs, indexed using 1-byte fingerprints (Pearson hashes) that speed locating a given key. Leaf modifications are accelerated using zero-copy updates.

Prerequisites

Libpmemobj-cpp package is required.

csmap

A persistent, concurrent and sorted engine, backed by a skip list. It is disabled by default. It can be enabled in CMake using the ENGINE_CSMAP option (requires C++14 support).

All methods of csmap are thread safe. Put, get, count_* and get_* scale with the number of threads. Remove method is currently implemented to take a global lock - it blocks all other threads.

Configuration

  • path -- Path to the database file
    • type: string
  • force_create -- If 0, pmemkv opens the file specified by 'path', otherwise it creates the file
    • type: uint64_t
    • default value: 0
  • size -- Only needed when force_create is not 0, specifies size of the database [in bytes]
    • type: uint64_t

stree

A persistent, single-threaded and sorted engine, backed by a B+ tree. It is disabled by default. It can be enabled in CMake using the ENGINE_STREE option.

Configuration

  • path -- Path to the database file
    • type: string
  • force_create -- If 0, pmemkv opens file specified by 'path', otherwise it creates it
    • type: uint64_t
    • default value: 0
  • size -- Only needed when force_create is not 0, specifies size of the database [in bytes]
    • type: uint64_t

Internals

(TBD)

Prerequisites

Libpmemobj-cpp package is required.

caching

This engine is using a sub engine from the list above to cache requests to external Redis or Memcached server. It is disabled by default. It can be enabled in CMake using the ENGINE_CACHING option.

Configuration

Caching engine itself requires server connection settings. Part of the config required for the sub engine should be relevant to chosen engine.

  • host -- Server's IP
    • type: string
  • port -- Server's port
    • type: int64_t
  • attempts -- Number of connection attempts
    • type: int64_t
  • ttl -- Time to live [in seconds]
    • type: int64_t
    • default value: 0
  • remote_type -- Server's type (Redis or Memcached)
    • type: string
  • remote_user -- Connection's user
    • type: string
  • remote_pwd -- User's password
    • type: string
  • remote_url -- Remote (server's) URL
    • type: string
  • subengine -- Config object for sub engine with its required settings
    • type: object

Internals

(TBD)

Prerequisites

Memcached and libacl (see here for installation guide) packages are required.

Related Work

pmse

tree3 has a lot in common with pmse -- both implementations rely on PMDK internally, although they expose different APIs externally. Both pmse and tree3 are based on a B+ tree implementation. The biggest difference is that the pmse tree keeps inner and leaf nodes in persistent memory, where tree3 keeps inner nodes in DRAM and leaf nodes in persistent memory. (This means that tree3 has to recover all inner nodes when the engine is started)

FPTree

This research paper describes a hybrid DRAM/NVM tree design (similar to the tree3 storage engine) but this paper doesn't provide any code, and omits certain important implementation details.

Beyond providing a clean-room implementation, the design of tree3 differs from FPTree in several important areas:

  1. tree3 is written using PMDK C++ bindings, which exerts influence on its design and implementation. tree3 uses generic PMDK transactions (i.e. transaction::run() closures), there is no need for micro-logging structures as described in the FPTree paper to make internal delete and split operations safe. tree3 also adjusts sizes of data structures (to fit PMDK primitive types) for best cache-line optimization.

  2. FPTree does not specify a hash method implementation, where tree3 uses a Pearson hash (RFC 3074).

  3. Within its persistent leaves, FPTree uses an array of key hashes with a separate visibility bitmap to track what hash slots are occupied. tree3 takes a different approach and uses key hashes themselves to track visibility. This relies on a specially modified Pearson hash function, where a hash value of zero always indicates the slot is unused. This optimization eliminates the cost of using and maintaining visibility bitmaps as well as cramming more hashes into a single cache-line, and affects the implementation of every primitive operation in the tree.

  4. tree3 caches key hashes in DRAM (in addition to storing these as part of the persistent leaf). This speeds leaf operations, especially with slower media, for what seems like an acceptable rise in DRAM usage.

  5. Within its persistent leaves, tree3 combines hash, key and value into a single slot type (KVSlot). This leads to improved leaf split performance and reduced write amplification, since splitting can be performed by swapping pointers to slots without copying any key or value data stored in the slots. KVSlot internally stores key and value to a single persistent buffer, which minimizes the number of persistent allocations and improves storage efficiency with larger keys and values.

cpp_map

Use of PMDK C++ bindings by tree3 was lifted from this example program. Many thanks to @tomaszkapela for providing a great example to follow!