ssmem
ssmem is a fast and simple object-based memory allocator with epoch-based garbage collection. In other words, the freed memory is not reused by the allocator until it is certain that no other threads might be accessing that memory. ssmem can for example be used in lock-free data structures, where the removal of an element from the structure might happen while other threads are holding a reference to the same element.
You can create multiple allocators with ssmem, however, each allocator can handle a single object size. For example, you can create an ssmem allocator to always allocate and release 64 bytes memory chunks.
- Website : http://lpd.epfl.ch/site/ascylib
- Author : Vasileios Trigonakis vasileios.trigonakis@epfl.ch
- Related Publications: ssmem was developed for: Asynchronized Concurrency: The Secret to Scaling Concurrent Search Data Structures, Tudor David, Rachid Guerraoui, Vasileios Trigonakis (alphabetical order), ASPLOS '15
Installation
Execute make
in the ssmem base folder.
This should compile libssmem.a
and ssmem_test
.
ssmem_test
contains a few tests for verifying the correctness and testing the performance of ssmem.
Execute: ssmem_test -h
for the available options.
If you want to customize the installation of ssmem, you can add a custom configuration for a platform in the Makefile
.
You can also compile with make VERSION=DEBUG
to generate a debug build of ssmem.
Using ssmem
To use ssmem you need to include ssmem.h
and link with -lssmem
.
ssmem.h
contains the interface of ssmem.
In short, you can:
ssmem_alloc_init
, orssmem_alloc_init_fs_size
to initialize and allocatorssmem_term
to terminate ssmemssmem_alloc
to allocate memoryssmem_free
to free memory (when it is safe, within the allocator)ssmem_release
to free memory (when it is safe, to the system -- withfree
)
Refer to ssmem.h
for more details and operations.
Memory reclaimation
Each thread holds a version/timestamp number. By default, this timestamp is incremented every time the thread takes some steps with ssmem (i.e., either allocates, or frees memory). ssmem batches multiple frees together (in a free set) and once the set is full, it tries to garbage collect memory. The free set is timestamped with a vector clock (an array of the timestamps from all threads). The garbage collection (reclamation of the freed memory) decides if a free set is safe to release (i.e., whether all threads had made progress since it was timestamped) and if it is, it makes the memory of this (and of "older") free sets available to the allocator.
Nevertheless, in certain cases, allocating and/or freeing memory might not be the correct place to increase the local timestamp. In a sense, incrementing the timestamp is some sort of barrier: by incrementing the timestamp the thread indicates that it does not hold any earlier references to ssmem-allocated memory. Accordingly, if the software allocating and/or freeing memory is not the correct "barrier" place for an application, ssmem includes a flag (in ssmem.h
) to control when the timestamp is incremented.
In brief, the flag SSMEM_TS_INCR_ON
can takes the following values (if the value is changed, recompilation is required):
SSMEM_TS_INCR_ON_NONE
: the timestamp is never automatically incremented. Applications should manually includeSSMEM_SAFE_TO_RECLAIM()
calls at points when it is safe (for other threads) to reclaim memory,SSMEM_TS_INCR_ON_BOTH
: the timestamp is incremented both upon allocating and freeing memory,SSMEM_TS_INCR_ON_ALLOC
: the timestamp is incremented only upon allocating memory,SSMEM_TS_INCR_ON_FREE
: the timestamp is incremented only upon freeing memory.