11 unstable releases (3 breaking)

0.4.3 Aug 9, 2020
0.4.2 Apr 24, 2020
0.3.1 Apr 23, 2020
0.3.0 Mar 10, 2020
0.1.0 Nov 26, 2019

#262 in Memory management

Download history 1543/week @ 2023-11-23 1636/week @ 2023-11-30 1720/week @ 2023-12-07 1332/week @ 2023-12-14 688/week @ 2023-12-21 314/week @ 2023-12-28 896/week @ 2024-01-04 2184/week @ 2024-01-11 4981/week @ 2024-01-18 5808/week @ 2024-01-25 6042/week @ 2024-02-01 4160/week @ 2024-02-08 8197/week @ 2024-02-15 9903/week @ 2024-02-22 12810/week @ 2024-02-29 2939/week @ 2024-03-07

34,853 downloads per month
Used in 9 crates (8 directly)

MPL-2.0+

69KB
1.5K SLoC

refpool

A reimplementation of Rust's std::boxed::Box and std::rc::Rc which uses a pool of reusable memory to speed up reallocation.

Is It Fast?

It's about twice as fast as the system allocator on Linux systems, and six times as fast on Windows systems, when the pool is non-empty. For certain data types, gains can be even higher.

Documentation

Licence

Copyright 2019 Bodil Stokke

This software is subject to the terms of the Mozilla Public License, v. 2.0. If a copy of the MPL was not distributed with this file, You can obtain one at http://mozilla.org/MPL/2.0/.

Code of Conduct

Please note that this project is released with a Contributor Code of Conduct. By participating in this project you agree to abide by its terms.


lib.rs:

A reimplementation of std::boxed::Box and std::rc::Rc which uses a pool of reusable memory to speed up reallocation.

Prerequisites

In order to initialise a type to its default value from the memory pool using PoolBox::default() or PoolRef::default(), it needs to implement PoolDefault.

If you want to be able to use PoolRef::make_mut(), it also needs to implement PoolClone.

For constructing values using PoolRef::new(), there's no requirement.

There are implementations for PoolDefault and PoolClone for most primitive types and a good selection of std's data types, and you can easily provide default implementations for your own types by implementing the marker trait PoolDefaultImpl. You can also implement your own if you have data structures whose memory doesn't need to be fully intitialised at construction time, which can give you a slight performance boost. (This optimisation is why PoolDefault and PoolClone exist as distinct traits, otherwise Default and Clone would have sufficed.)

Usage

You create new values by calling PoolRef::default(pool) or PoolRef::new(pool, value). This will use memory from the pool if available, falling back to a normal heap allocation if the pool is empty. When the last PoolRef referencing the value is dropped, its allocated memory is returned to the pool.

Differences from Box and Rc

PoolBox is API compatible with Box and PoolRef with Rc, with the following exceptions:

Thread Safety

Pool is strictly thread local, ie. it does not implement Sync and it will fail in appalling ways if you still somehow manage to access it from two different threads. There is no equivalent of Arc because adding thread safety to the pool turns out to degrade performance sufficiently that the pool is no longer providing a significant performance benefit even with the slowest system allocators you're likely to come across in the wild (by which I mean Windows).

Performance

You can expect Pool to always outperform the system allocator, though the performance gains will vary between platforms. Preliminary benchmarks show it's approximately twice as fast on Linux, and 5-6 times as fast on Windows. Custom allocators like jemalloc may yield even less benefit, but it's very unlikely you'll find an allocator that can outperform the pool.

You can expect bigger performance gains from data types with beneficial PoolDefault and PoolClone implementations, "beneficial" in this case meaning cases where you can leave most of the allocated memory uninitialised. sized_chunks::Chunk, which allocates 528 bytes on 64-bit platforms but only needs to initialise 16 of them for PoolDefault, would be a good example of this.

Example

// Create a pool of `usize` with a max size of 1 (for argument's sake).
let mut pool: Pool<usize> = Pool::new(1);

{
    // Create a reference handle to a usize initialised to 0.
    // The pool starts out empty, so this triggers a normal heap alloc.
    let value_ref = PoolRef::default(&mut pool);
    assert_eq!(0, *value_ref); // You can deref it just like `Rc`.
} // `value_ref` is dropped here, and its heap memory goes on the pool.

// Check that we do indeed have one allocation in the pool now.
assert_eq!(1, pool.get_pool_size());

// Create another reference and initialise it to 31337, a good round number.
// This will reuse `value_ref`'s memory.
let another_value_ref = PoolRef::new(&mut pool, 31337);
assert_eq!(31337, *another_value_ref);

// Check that the pool is again empty after we reused the previous memory.
assert_eq!(0, pool.get_pool_size());

Feature Flags

There's one feature flag available, default_impl, which requires a nightly rustc because it leans on the min_specialization language feature, which removes the PoolDefaultImpl trait and instead provides a default overridable implementation for PoolClone and PoolDefault for any type that implements Clone and Default. PoolDefaultImpl is an unfortunate hack to get around the current absence of specialisation in stable rustc.

No runtime deps