rsbmalloc
| Crates.io | rsbmalloc |
| lib.rs | rsbmalloc |
| version | 0.4.4 |
| source | src |
| created_at | 2023-04-03 14:51:10.500646 |
| updated_at | 2024-04-23 15:34:52.503467 |
| description | A simple but fast binned allocator for Rust. Single-threaded for no_std, multi-threaded otherwise |
| homepage | https://github.com/AWBroch/rsbmalloc |
| repository | https://github.com/AWBroch/rsbmalloc |
| max_upload_size | |
| id | 829146 |
| size | 28,031 |
(AWBroch)
documentation
README
rsbmalloc
A binned allocator for Rust. It’s quite simple, but reasonably fast single and multi-threaded. Single-threaded, it generally similar to the built-in allocator, sometimes faster, but sometimes with higher memory usage. Multi-threaded, it ranges from similar speed to quite a bit slower. It’s pure Rust, so it should work smoothly on any platform that provides standard mmap and munmap functions, and also Windows (though Windows support isn’t tested).
Relies exclusively on thread-local caches for multi-threaded support. 4 times the number of cpus are created on the first allocation and no more are created after that, so each ‘thread-local’ cache is fully thread-safe in case it is reused between threads.
rsbmalloc is entirely a binned allocator, with bins ranging from 4 bytes to 16 KiB (some ARM pages sizes are 16 KiB). If an allocation is larger than 16 KiB, it gets counted as a large allocation and goes straight to mmap and munmap. So, when freed in Rust, it gets munmap-ed. Bins, however, are allocated a page at a time as necessary and are never released back to the OS. Freed slots just act as a linked list that can be reused by the same thread (or another thread that scores the same thread cache).
It implements the GlobalAllocator trait, and comes with a single-threaded no_std version. The no_std version still requires a libc with mmap and munmap or Windows, but it doesn’t depend on the Rust standard library. Note that the no_std version is still thead-safe, it just doesn’t use the thread-local caches, so it’s a lot slower because it relies of spinlocks when operating multi-threaded. On the other hand, it uses less memory and would be a similar speed if there’s no lock contention. Once the allocator-api is stable, it should be a fairly easy port to that.
rsbmallocc provides a slightly slower C interface to rsbmalloc, including standard names (malloc, free) and prefixed names (rsbmalloc, rsbfree).
rsbmalloc also exposes the page-only allocator it uses under the hood.
A Broch Web Solutions project.
Check out the blog post for more info.