A simple no_std allocator written in Rust that manages memory in fixed-size chunks/blocks. Useful for basic no_std
binaries where you want to manage a heap of a few megabytes without complex features such as paging/page table
management. Instead, this allocator gets a fixed/static memory region and allocates memory from there. This memory
region can be contained inside the executable file that uses this allocator. See examples down below.
⚠ Other allocators with different properties (for example better memory utilization but less performance) do exist. The README of the repository contains a section that discusses how this allocator relates to other existing allocators on <crates.io>. ⚠
- ✅
no_stdallocator with test coverage - ✅ uses static memory as backing storage (no paging/page table manipulations)
- ✅ allocation strategy is a combination of next-fit and best-fit
- ✅ reasonable fast with low code complexity
- ✅ const compatibility (no runtime
init()required) - ✅ efficient in scenarios where heap is a few dozens megabytes in size
- ✅ user-friendly API
The inner and low-level ChunkAllocator can be used as #[global_allocator] with the synchronized wrapper type
GlobalChunkAllocator. Both can be used with the allocator_api feature. The latter enables the usage in several
types of the Rust standard library, such as Vec::new_in or BTreeMap::new_in. This is primarily interesting for
testing but may also enable other interesting use-cases.
The focus is on const compatibility. The allocator and the backing memory can get initialized during compile time
and need no runtime init() call or similar. This means that if the compiler accepts it then the allocation will
also work during runtime. However, you can also create allocator objects during runtime.
The inner and low-level ChunkAllocator is a chunk allocator or also called fixed-size block allocator. It uses a
mixture of the strategies next-fit and a best-fit. It tries to use the smallest gap for an allocation request to
prevent fragmentation but this is no guarantee. Each allocation is a trade-off between a low allocation time and
preventing fragmentation. The default chunk size is 256 bytes but this can be changed as compile time const generic.
Having a fixed-size block allocator enables an easy bookkeeping algorithm through a bitmap but has as consequence that
small allocations, such as 64 byte will take at least one chunk/block of the chosen block size.
This project originates from my Diplom thesis project. Since I originally had lots of struggles to create this (my first ever allocator), I outsourced it for better testability and to share my knowledge and findings with others in the hope that someone can learn from it in any way.
#![feature(const_mut_refs)]
#![feature(allocator_api)]
use simple_chunk_allocator::{heap, heap_bitmap, GlobalChunkAllocator, PageAligned};
// The macros help to get a correctly sized arrays types.
// I page-align them for better caching and to improve the availability of
// page-aligned addresses.
/// Backing storage for heap (1Mib). (read+write) static memory in final executable.
///
/// heap!: first argument is chunk amount, second argument is size of each chunk.
/// If no arguments are provided it falls back to defaults.
/// Example: `heap!(chunks=16, chunksize=256)`.
static mut HEAP: PageAligned<[u8; 1048576]> = heap!();
/// Backing storage for heap bookkeeping bitmap. (read+write) static memory in final executable.
///
/// heap_bitmap!: first argument is amount of chunks.
/// If no argument is provided it falls back to a default.
/// Example: `heap_bitmap!(chunks=16)`.
static mut HEAP_BITMAP: PageAligned<[u8; 512]> = heap_bitmap!();
// please make sure that the backing memory is at least CHUNK_SIZE aligned; better page-aligned
#[global_allocator]
static ALLOCATOR: GlobalChunkAllocator =
unsafe { GlobalChunkAllocator::new(HEAP.deref_mut_const(), HEAP_BITMAP.deref_mut_const()) };
fn main() {
// at this point, the allocator already got used a bit by the Rust runtime that executes
// before main() gets called. This is not the case if a `no_std` binary gets produced.
let old_usage = ALLOCATOR.usage();
let mut vec = Vec::new();
vec.push(1);
vec.push(2);
vec.push(3);
assert!(ALLOCATOR.usage() > old_usage);
// use "allocator_api"-feature. You can use this if "ALLOCATOR" is not registered as
// the global allocator. Otherwise, it is already the default.
let _boxed = Box::new_in([1, 2, 3], ALLOCATOR.allocator_api_glue());
}This is an excerpt. The code can be found in the GitHub repository in freestanding-linux-example.
static mut HEAP: PageAligned<[u8; 256]> = heap!(chunks = 16, chunksize = 16);
static mut HEAP_BITMAP: PageAligned<[u8; 2]> = heap_bitmap!(chunks = 16);
// please make sure that the backing memory is at least CHUNK_SIZE aligned; better page-aligned
#[global_allocator]
static ALLOCATOR: GlobalChunkAllocator<16> =
unsafe { GlobalChunkAllocator::<16>::new(HEAP.deref_mut_const(), HEAP_BITMAP.deref_mut_const()) };
/// Referenced as entry by linker argument. Entry into the code.
#[no_mangle]
fn start() -> ! {
write!(StdoutWriter, "Hello :)\n").unwrap();
let mut vec = Vec::new();
(0..10).for_each(|x| vec.push(x));
write!(StdoutWriter, "vec: {:#?}\n", vec).unwrap();
exit();
}This crate only builds with the nightly version of Rust because it uses many nightly-only features. I developed it
with version 1.61.0-nightly (2022-03-05). Older nightly versions might work. So far there is no stable Rust
compiler version that compiles this.
The default CHUNK_SIZE is 256 bytes. It is a tradeoff between performance and efficient memory usage.
I executed my example bench in release mode on an Intel i7-1165G7 CPU and a heap of 160MB to get the results listed
below. I used RUSTFLAGS="-C target-cpu=native" cargo run --release --example bench to excute the benchmark with
maximum performance. The benchmark simulates a heavy usage of the heap in a single-threaded program with many random
allocations and deallocations. The benchmark stops when the heap is close to 100%. The allocations vary in their alignment.
The table below shows the results of this benchmark as number of clock cycles.
Info: Since I measured those values, I slightly changed the benchmark.
| Chunk Size | # Chunks | # allocations | # deallocations | median | average | min | max |
|---|---|---|---|---|---|---|---|
| 128 | 1310720 | 68148 | 47915 | 955 | 1001 | 126 | 57989 |
| 256 [DEFAULT] | 655360 | 71842 | 51744 | 592 | 619 | 121 | 53578 |
| 512 | 327680 | 66672 | 46858 | 373 | 401 | 111 | 54403 |
The results vary slightly because each run gets influenced by some randomness. One can see that the performance gets slower with a growing number of chunks. Increasing the chunk size reduces the size of the bookkeeping bitmap which accelerates the lookup. However, a smaller chunk size occupies less heap when only very small allocations are required.
Note that performance is better than listed above when the heap is used less frequently and does not run full.
Update November 2022: I recently found this new project and, from a first glance, I recommend to use this crate instead of mine for production usage. It has impressive performance and heap utilization at the costs of more complicated code. The repository includes interesting performance numbers from galloc, simple-chunk-allocator (this crate), and linked-list-allocator.
Update November 2022: I wrote this paragraph before I found out about galloc. I left it unchanged.
The linked-list-allocator is among the few other well-suited and maintained general-purpose no-std allocator I could find on crates.io.
Advantages of my chunk allocator:
- much faster median allocation time
- much faster average allocation time [ONLY IF HEAP IS NOT CLOSE TO BEEING FULL]
- optimized realloc in certain cases (almost a no-op in some situations)
- uses relatively easy algorithm (but needs dedicated heap and book-keeping backing storage)
Advantages of linked-list-allocator:
- better memory utilization (less fragmentation)
- better worst-case allocation time in most test runs
- better average allocation time [ONLY IF HEAP IS CLOSE TO BEEING FULL]
- only needs a single chunk of memory and manages the heap with the backing-memory itself
Benchmark Comparision:
I ran $ cargo run --example bench --release against both allocators and obtained the following results. The benchmark
performs random allocations of different sizes and alignments and also deallocates some of the older allocations. Over
time, the heap becomes full, which is why the number of successful allocations has a higher delta to the attempted
allocations.
Runtime: 1s (most time lot's of heap available)
RESULTS OF BENCHMARK: Chunk Allocator
53360 allocations, 16211 successful_allocations, 37149 deallocations
median= 878 ticks, average= 1037 ticks, min= 158 ticks, max= 7178941 ticks
RESULTS OF BENCHMARK: Linked List Allocator
31627 allocations, 9374 successful_allocations, 22253 deallocations
median= 18582 ticks, average= 44524 ticks, min= 71 ticks, max=44126026 ticks
We see that as long as most allocations are done on a heap with lots of space available, the chunk allocator is faster in median and average performance.
Runtime: 10s (most time heap almost full)
RESULTS OF BENCHMARK: Chunk Allocator
74909 allocations, 23753 successful_allocations, 51156 deallocations
median= 961 ticks, average=273362 ticks, min= 167 ticks, max=53330953 ticks
RESULTS OF BENCHMARK: Linked List Allocator
81884 allocations, 24792 successful_allocations, 57092 deallocations
median=100196 ticks, average=179495 ticks, min= 69 ticks, max=43937820 ticks
We see that when the heap is almost full, the chunk allocator has a faster median performance but a worse worst-case allocation time. The linked list allocator performs better on average (but not on median) when it is close to beeing full.