/rpmalloc

Public domain cross platform lock free thread caching 32-byte aligned memory allocator implemented in C

Primary LanguageCOtherNOASSERTION

rpmalloc - Rampant Pixels Memory Allocator

This library provides a public domain cross platform lock free thread caching 32-byte aligned memory allocator implemented in C. The latest source code is always available at https://github.com/rampantpixels/rpmalloc

Platforms currently supported:

  • Windows
  • MacOS
  • iOS
  • Linux
  • Android

The code should be easily portable to any platform with atomic operations and an mmap-style virtual memory management API. The API used to map/unmap memory pages can be configured in runtime to a custom implementation and mapping granularity/size.

This library is put in the public domain; you can redistribute it and/or modify it without any restrictions. Or, if you choose, you can use it under the MIT license.

Please consider our Patreon to support our work - https://www.patreon.com/rampantpixels

Created by Mattias Jansson (@maniccoder) / Rampant Pixels - http://www.rampantpixels.com

Performance

We believe rpmalloc is faster than most popular memory allocators like tcmalloc, hoard, ptmalloc3 and others without causing extra allocated memory overhead in the thread caches compared to these allocators. We also believe the implementation to be easier to read and modify compared to these allocators, as it is a single source file of ~2000 lines of C code. All allocations have a natural 32-byte alignment.

Contained in a parallel repository is a benchmark utility that performs interleaved allocations (both aligned to 8 or 16 bytes, and unaligned) and deallocations (both in-thread and cross-thread) in multiple threads. It measures number of memory operations performed per CPU second, as well as memory overhead by comparing the virtual memory mapped with the number of bytes requested in allocation calls. The setup of number of thread, cross-thread deallocation rate and allocation size limits is configured by command line arguments.

https://github.com/rampantpixels/rpmalloc-benchmark

Below is an example performance comparison chart of rpmalloc and other popular allocator implementations, with default configurations used.

Ubuntu 16.10, random [16, 8000] bytes, 8 cores

The benchmark producing these numbers were run on an Ubuntu 16.10 machine with 8 logical cores (4 physical, HT). The actual numbers are not to be interpreted as absolute performance figures, but rather as relative comparisons between the different allocators. For additional benchmark results, see the BENCHMARKS file.

Configuration of the thread and global caches can be important depending on your use pattern. See CACHE for a case study and some comments/guidelines.

Using

The easiest way to use the library is simply adding rpmalloc.[h|c] to your project and compile them along with your sources. This contains only the rpmalloc specific entry points and does not provide internal hooks to process and/or thread creation at the moment. You are required to call these functions from your own code in order to initialize and finalize the allocator in your process and threads:

rpmalloc_initialize : Call at process start to initialize the allocator

rpmalloc_initialize_config : Optional entry point to call at process start to initialize the allocator with a custom memory mapping backend, memory page size and mapping granularity.

rpmalloc_finalize: Call at process exit to finalize the allocator

rpmalloc_thread_initialize: Call at each thread start to initialize the thread local data for the allocator

rpmalloc_thread_finalize: Call at each thread exit to finalize and release thread cache back to global cache

rpmalloc_config: Get the current runtime configuration of the allocator

Then simply use the rpmalloc/rpfree and the other malloc style replacement functions. Remember all allocations are 16-byte aligned, so no need to call the explicit rpmemalign/rpaligned_alloc/rpposix_memalign functions unless you need greater alignment, they are simply wrappers to make it easier to replace in existing code.

If you wish to override the standard library malloc family of functions and have automatic initialization/finalization of process and threads, also include the malloc.c file in your project. The automatic init/fini is only implemented for Linux and macOS targets. The list of libc entry points replaced may not be complete, use libc replacement only as a convenience for testing the library on an existing code base, not a final solution.

Building

To compile as a static library run the configure python script which generates a Ninja build script, then build using ninja. The ninja build produces two static libraries, one named rpmalloc and one named rpmallocwrap, where the latter includes the libc entry point overrides.

The configure + ninja build also produces two shared object/dynamic libraries. The rpmallocwrap shared library can be used with LD_PRELOAD/DYLD_INSERT_LIBRARIES to inject in a preexisting binary, replacing any malloc/free family of function calls. This is only implemented for Linux and macOS targets. The list of libc entry points replaced may not be complete, use preloading as a convenience for testing the library on an existing binary, not a final solution.

The latest stable release is available in the master branch. For latest development code, use the develop branch.

Cache configuration options

Free memory pages are cached both per thread and in a global cache for all threads. The size of the thread caches is determined by an adaptive scheme where each cache is limited by a percentage of the maximum allocation count of the corresponding size class. The size of the global caches is determined by a multiple of the maximum of all thread caches. The factors controlling the cache sizes can be set by editing the individual defines in the rpmalloc.c source file for fine tuned control.

ENABLE_UNLIMITED_CACHE: By default defined to 0, set to 1 to make all caches infinite, i.e never release spans to global cache unless thread finishes and never unmap memory pages back to the OS. Highest performance but largest memory overhead.

ENABLE_UNLIMITED_GLOBAL_CACHE: By default defined to 0, set to 1 to make global caches infinite, i.e never unmap memory pages back to the OS.

ENABLE_UNLIMITED_THREAD_CACHE: By default defined to 0, set to 1 to make thread caches infinite, i.e never release spans to global cache unless thread finishes.

ENABLE_GLOBAL_CACHE: By default defined to 1, enables the global cache shared between all threads. Set to 0 to disable the global cache and directly unmap pages evicted from the thread cache.

ENABLE_THREAD_CACHE: By default defined to 1, enables the per-thread cache. Set to 0 to disable the thread cache and directly unmap pages no longer in use (also disables the global cache).

Other configuration options

Detailed statistics are available if ENABLE_STATISTICS is defined to 1 (default is 0, or disabled), either on compile command line or by setting the value in rpmalloc.c. This will cause a slight overhead in runtime to collect statistics for each memory operation, and will also add 4 bytes overhead per allocation to track sizes.

Integer safety checks on all calls are enabled if ENABLE_VALIDATE_ARGS is defined to 1 (default is 0, or disabled), either on compile command line or by setting the value in rpmalloc.c. If enabled, size arguments to the global entry points are verified not to cause integer overflows in calculations.

Asserts are enabled if ENABLE_ASSERTS is defined to 1 (default is 0, or disabled), either on compile command line or by setting the value in rpmalloc.c.

Overwrite and underwrite guards are enabled if ENABLE_GUARDS is defined to 1 (default is 0, or disabled), either on compile command line or by settings the value in rpmalloc.c. This will introduce up to 64 byte overhead on each allocation to store magic numbers, which will be verified when freeing the memory block. The actual overhead is dependent on the requested size compared to size class limits.

Huge pages

The allocator has support for huge/large pages on Windows, Linux and MacOS. To enable it, pass a non-zero value in the config value enable_huge_pages when initializing the allocator with rpmalloc_initialize_config. If the system does not support huge pages it will be automatically disabled. You can query the status by looking at enable_huge_pages in the config returned from a call to rpmalloc_config after initialization is done.

Quick overview

The allocator is similar in spirit to tcmalloc from the Google Performance Toolkit. It uses separate heaps for each thread and partitions memory blocks according to a preconfigured set of size classes, up to 2MiB. Larger blocks are mapped and unmapped directly. Allocations for different size classes will be served from different set of memory pages, each "span" of pages is dedicated to one size class. Spans of pages can flow between threads when the thread cache overflows and are released to a global cache, or when the thread ends. Unlike tcmalloc, single blocks do not flow between threads, only entire spans of pages.

Implementation details

The allocator is based on a fixed but configurable page alignment (defaults to 64KiB) and 32 byte block alignment, where all runs of memory pages (spans) are mapped to this alignment boundary. On Windows this is automatically guaranteed up to 64KiB by the VirtualAlloc granularity, and on mmap systems it is achieved by oversizing the mapping and aligning the returned virtual memory address to the required boundaries. By aligning to a fixed size the free operation can locate the header of the memory span without having to do a table lookup (as tcmalloc does) by simply masking out the low bits of the address (for 64KiB this would be the low 16 bits).

Memory blocks are divided into three categories. For 64KiB span size/alignment the small blocks are [32, 2016] bytes, medium blocks (2016, 32720] bytes, and large blocks (32720, 2097120] bytes. The three categories are further divided in size classes. If the span size is changed, the small block classes remain but medium blocks go from (2016, span size] bytes.

Small blocks have a size class granularity of 32 bytes each in 63 buckets. Medium blocks have a granularity of 512 bytes, 60 buckets (default). Large blocks have a the same granularity as the configured span size (default 64KiB). All allocations are fitted to these size class boundaries (an allocation of 42 bytes will allocate a block of 64 bytes). Each small and medium size class has an associated span (meaning a contiguous set of memory pages) configuration describing how many pages the size class will allocate each time the cache is empty and a new allocation is requested.

Spans for small and medium blocks are cached in four levels to avoid calls to map/unmap memory pages. The first level is a per thread single active span for each size class. The second level is a per thread list of partially free spans for each size class. The third level is a per thread list of free spans. The fourth level is a global list of free spans.

Each span for a small and medium size class keeps track of how many blocks are allocated/free, as well as a list of which blocks that are free for allocation. To avoid locks, each span is completely owned by the allocating thread, and all cross-thread deallocations will be deferred to the owner thread.

Large blocks, or super spans, are cached in two levels. The first level is a per thread list of free super spans. The second level is a global list of free super spans.

Memory mapping

By default the allocator uses OS APIs to map virtual memory pages as needed, either VirtualAlloc on Windows or mmap on POSIX systems. If you want to use your own custom memory mapping provider you can use rpmalloc_initialize_config and pass function pointers to map and unmap virtual memory. These function should reserve and free the requested number of bytes.

The returned memory address from the memory map function MUST be aligned to the memory page size and the memory span size (which ever is larger), both of which is configurable. Either provide the page and span sizes during initialization using rpmalloc_initialize_config, or use rpmalloc_config to find the required alignment which is equal to the maximum of page and span size. The span size MUST be a power of two in [4096, 262144] range, and be a multiple or divisor of the memory page size.

Memory mapping requests are always done in multiples of the memory page size. You can specify a custom page size when initializing rpmalloc with rpmalloc_initialize_config, or pass 0 to let rpmalloc determine the system memory page size using OS APIs. The page size MUST be a power of two.

To reduce system call overhead, memory spans are mapped in batches controlled by the span_map_count configuration variable (which defaults to the DEFAULT_SPAN_MAP_COUNT value if 0, which in turn is sized according to the cache configuration define, defaulting to 32). If the memory page size is larger than the span size, the number of spans to map in a single call will be adjusted to guarantee a multiple of the page size, and the spans will be kept mapped until the entire span range can be unmapped in one call (to avoid trying to unmap partial pages).

Span breaking

Super spans (spans a multiple > 1 of the span size) can be subdivided into smaller spans to fulfull a need to map a new span of memory. By default the allocator will greedily grab and break any larger span from the available caches before mapping new virtual memory. However, spans can currently not be glued together to form larger super spans again. Subspans can traverse the cache and be used by different threads individually.

A span that is a subspan of a larger super span can be individually decommitted to reduce physical memory pressure when the span is evicted from caches and scheduled to be unmapped. The entire original super span will keep track of the subspans it is broken up into, and when the entire range is decommitted tha super span will be unmapped. This allows platforms like Windows that require the entire virtual memory range that was mapped in a call to VirtualAlloc to be unmapped in one call to VirtualFree, while still decommitting individual pages in subspans (if the page size is smaller than the span size).

If you use a custom memory map/unmap function you need to take this into account by looking at the release parameter given to the memory_unmap function. It is set to 0 for decommitting invididual pages and the total super span byte size for finally releasing the entire super span memory range.

Memory guards

If you define the ENABLE_GUARDS to 1, all memory allocations will be padded with extra guard areas before and after the memory block (while still honoring the requested alignment). These dead zones will be filled with a pattern and checked when the block is freed. If the patterns are not intact the callback set in initialization config is called, or if not set an assert is fired.

Note that the end of the memory block in this case is defined by the total usable size of the block as returned by rpmalloc_usable_size, which can be larger than the size passed to allocation request due to size class buckets.

Memory fragmentation

There is no memory fragmentation by the allocator in the sense that it will not leave unallocated and unusable "holes" in the memory pages by calls to allocate and free blocks of different sizes. This is due to the fact that the memory pages allocated for each size class is split up in perfectly aligned blocks which are not reused for a request of a different size. The block freed by a call to rpfree will always be immediately available for an allocation request within the same size class.

However, there is memory fragmentation in the meaning that a request for x bytes followed by a request of y bytes where x and y are at least one size class different in size will return blocks that are at least one memory page apart in virtual address space. Only blocks of the same size will potentially be within the same memory page span.

Unlike the similar tcmalloc where the linked list of individual blocks leads to back-to-back allocations of the same block size will spread across a different span of memory pages each time (depending on free order), rpmalloc keeps an "active span" for each size class. This leads to back-to-back allocations will most likely be served from within the same span of memory pages (unless the span runs out of free blocks). The rpmalloc implementation will also use any "holes" in memory pages in semi-filled spans before using a completely free span.

Producer-consumer scenario

Compared to the tcmalloc implementation, rpmalloc does not suffer as much from a producer-consumer thread scenario where one thread allocates memory blocks and another thread frees the blocks. In tcmalloc the free blocks need to traverse both the thread cache of the thread doing the free operations as well as the global cache before being reused in the allocating thread. In rpmalloc the freed blocks will be reused as soon as the allocating thread needs to get new spans from the thread cache.

Best case scenarios

Threads that keep ownership of allocated memory blocks within the thread and free the blocks from the same thread will have optimal performance.

Threads that have allocation patterns where the difference in memory usage high and low water marks fit within the thread cache thresholds in the allocator will never touch the global cache except during thread init/fini and have optimal performance. Tweaking the cache limits can be done on a per-size-class basis.

Worst case scenarios

Since each thread cache maps spans of memory pages per size class, a thread that allocates just a few blocks of each size class (32, 64, ...) for many size classes will never fill each bucket, and thus map a lot of memory pages while only using a small fraction of the mapped memory. However, the wasted memory will always be less than 64KiB (or the configured span size) per size class. The cache for free spans will be reused by all size classes.

An application that has a producer-consumer scheme between threads where one thread performs all allocations and another frees all memory will have a sub-optimal performance due to blocks crossing thread boundaries will be freed in a two step process - first deferred to the allocating thread, then freed when that thread has need for more memory pages for the requested size. However, depending on the use case the performance overhead might be small.

Threads that perform a lot of allocations and deallocations in a pattern that have a large difference in high and low water marks, and that difference is larger than the thread cache size, will put a lot of contention on the global cache. What will happen is the thread cache will overflow on each low water mark causing pages to be released to the global cache, then underflow on high water mark causing pages to be re-acquired from the global cache. This can be mitigated by changing the MAX_SPAN_CACHE_DIVISOR define in the source code (at the cost of higher average memory overhead).

Caveats

Cross-thread deallocations are more costly than in-thread deallocations, since the spans are completely owned by the allocating thread. The free operation will be deferred using an atomic list operation and the actual free operation will be performed when the owner thread requires a new block of the corresponding size class.

VirtualAlloc has an internal granularity of 64KiB. However, mmap lacks this granularity control, and the implementation instead oversizes the memory mapping with configured span size to be able to always return a memory area with the required alignment. Since the extra memory pages are never touched this will not result in extra committed physical memory pages, but rather only increase virtual memory address space.

The free, realloc and usable size functions all require the passed pointer to be within the first 64KiB (or whatever you set the span size to) of the start of the memory block. You cannot pass in any pointer from the memory block address range.

All entry points assume the passed values are valid, for example passing an invalid pointer to free would most likely result in a segmentation fault. The library does not try to guard against errors.