lipaul / go-slab

slab allocator in go

go-slab - slab allocator in go

A slab allocator library in the Go Programming Language.

Who is this for

This library may be interesting to you if you wish to reduce garbage collection (e.g. stop-the-world GC) performance issues in your golang programs, allowing you to switch to explicit byte array memory management techniques.

This can be useful, for example, for long-running server programs that manage lots of in-memory data items, such as caches and databases.

Example usage

arena := NewArena(48,         // The smallest slab class "chunk size" is 48 bytes.
                  1024*1024,  // Each slab will be 1MB in size.
                  2,          // Power of 2 growth in "chunk sizes".
                  nil)        // Use default make([]byte) for slab memory.

var buf []byte

buf = arena.Alloc(64)         // Allocate 64 bytes.
  ... use the buf ...
arena.DecRef(buf)             // Release our ref-count when we're done with buf.

buf = arena.Alloc(1024)       // Allocate another 1K byte array.
  ... use the buf ...
arena.AddRef(buf)             // The buf's ref-count now goes to 2.
  ... use the buf some more ...
arena.DecRef(buf)
  ... still can use the buf since we still have 1 ref-count ...
arena.DecRef(buf)             // We shouldn't use the buf after this last DecRef(),
                              // as the library might recycle it for a future Alloc().

Design concepts

The byte arrays ([]byte) that are allocated by this library are reference-counted. When a byte array's reference count drops to 0, it will be placed onto a free-list for later re-use. This can reduce the need to ask the go runtime to allocate new memory and perhaps delay the need for a full stop-the-world GC.

The AddRef()/DecRef() functions use slice/capacity math instead of "large" additional tracking data structures (e.g., no extra hashtables) in order to reach the right ref-counter metadata.

This implementation also does not use any of go's "unsafe" capabilities, allowing it to remain relatively simple.

Memory is managed via a simple slab allocator algorithm. See: http://en.wikipedia.org/wiki/Slab_allocation

Each arena tracks one or more slabClass structs. Each slabClass manages a different "chunk size", where chunk sizes are computed using a simple "growth factor" (e.g., the "power of 2 growth" in the above example). Each slabClass also tracks zero or more slabs, where every slab tracked by a slabClass will all have the same chunk size. A slab manages a (usually large) continguous array of memory bytes (1MB from the above example), and the slab's memory is subdivided into many chunks of the same chunk size. All the chunks in a new slab are placed on a free-list that's part of the slabClass.

When Alloc() is invoked, the first "large enough" slabClass is found, and a chunk from the free-list is taken to service the allocation. If there are no more free chunks available in a slabClass, then a new slab (e.g., 1MB) is allocated, chunk'ified, and the request is processed as before.

Concurrency

The Arena returned from NewArena() is not concurrency safe. Please use your own locking.

Chainability

The []byte buf's can be chained via the SetNext()/GetNext() functions. This may be useful for developers wishing to reduce fragmentation when they have wildly varying byte array sizes.

For example, a server cache may need to manage many items whose sizes range from small to large (16 bytes to 1MB). Instead of invoking Arena.Alloc() on the exact item size, the developer may wish to consider slicing an item into many more smaller 4KB byte arrays.

For a 1MB item, for example, the application can instead invoke Arena.Alloc(4096) for 256 times and use the Arena.SetNext() function to chain those smaller 4KB buffers together. By slicing memory into uniform-sized, smaller-sized buffers, there may be less fragmentation and better overall re-use of slabs. Additionally, the last []byte buffer in the chain may be smaller than 4KB to not waste space.

Application specific slab memory allocator

The NewArena() function takes an optional malloc() callback function, which will be invoked whenever the arena needs more memory for a new slab. If the malloc() func is nil, the arena will default to using the builtin make([]byte, sizeNeeded).

An application-specific malloc() func can be useful for tracking and/or limiting the amount of slab memory that an Arena uses. It can be also used by advanced applications to supply mmap()'ed memory to an Arena.

Rules

• You need to invoke AddRef()/DecRef() with the exact same buf that you received from Alloc(), from the same arena.
• Don't call Alloc() with a size greater than the arena's slab size. e.g., if your slab size is 1MB, then Alloc(1024 * 1024 + 1) will fail.
• Careful with your ref-counting -- that's the fundamental tradeoff with now trying to avoid GC.
• Do not grow or append() on the slices returned by Alloc().
• Do not use cap() on slices returned by Alloc(), as that has information / abstraction "leakage" and should not be depended on.

LICENSE

Apache 2 license.

Testing

Unit test code coverage, as of version 0.0.0-42-g60296ca, is 99.4%.

TODO

• Currently, slabs that are allocated are never freed.
• Memory for one slabClass is never reassigned to another slabClass. Memory reassignment might be useful whenever data sizes of items in long-running systems change over time. For example, sessions in an online game may initially fit fine into a 1K slab class, but start getting larger than 1K as long time players acquire more inventory. Meanwhile, most of the slab memory is "stuck" in the 1K slab class when it's now needed in the 2K slab class. The chainability features of go-slab, of note, should also be considered in these cases.