5 releases

new 1.0.0-alpha.5 Jun 20, 2024
1.0.0-alpha.4 May 29, 2024
1.0.0-alpha.3 Feb 23, 2024
1.0.0-alpha.1 Dec 26, 2023

#44 in Concurrency

Download history 657/week @ 2024-03-03 675/week @ 2024-03-10 266/week @ 2024-03-17 331/week @ 2024-03-24 277/week @ 2024-03-31 188/week @ 2024-04-07 141/week @ 2024-04-14 220/week @ 2024-04-21 107/week @ 2024-04-28 207/week @ 2024-05-05 306/week @ 2024-05-12 393/week @ 2024-05-19 649/week @ 2024-05-26 468/week @ 2024-06-02 208/week @ 2024-06-09 629/week @ 2024-06-16

1,990 downloads per month
Used in 10 crates (5 directly)

MIT license

1.5MB
24K SLoC

hwlocality: Rust bindings for the hwloc library

MIT licensed Package on crates.io Documentation Continuous Integration Code coverage CII Best Practices Summary Requires rustc 1.71.0+

What is this?

To optimize programs for parallel performance on all hardware, you must accept that on many common platforms, symmetric multiprocessing is a lie. "CPUs" detected by the operating system often have inequal access to shared resources like caches, DRAM and I/O peripherals, sometimes even inequal specifications (as in Arm big.LITTLE, Apple Mx and Intel Adler Lake), and significant performance gains can be achieved by taking these facts into account in your code.

This is the latest maintained Rust binding to hwloc, a C library from Open MPI for detecting the hierarchical topology of modern architectures: NUMA memory nodes, sockets, shared data & instruction caches, cores, simultaneous multi threading, and more. Additionally, hwloc lets you pin threads to specific CPU cores and memory to specific NUMA nodes, which is a prerequisite to perform topology-aware program optimizations.

hwlocality is based on and still shares some code and design with the previous, now-unmaintained attempts to write Rust hwloc bindings at Ichbinjoe/hwloc2-rs and daschl/hwloc-rs. However, it does not aim for API compatibility with them. Indeed, many changes have been made with respect to hwloc2-rs in the aim of improving ergonomics, performance, and removing avenues for Undefined Behaviour like assuming pointers are non-null or union fields are valid when nobody tells you they will always be.

Prerequisites

hwlocality is compatible with libhwloc v2.0 and later. You can install a suitable version of libhwloc in two different ways:

  1. If your package manager of choice provides a reasonably recent libhwloc package, then you can install it along with the associated development package (typically called libhwloc-dev or libhwloc-devel). This is the recommended way to do things because it will greatly speed up your hwlocality (re)builds and allow you to easily keep libhwloc up to date along with the rest of your development environment.
  2. If you cannot use the above method for any reason, then hwlocality can alternatively download and build its own copy of libhwloc. To use such an internal build, please enable the vendored Cargo feature. In addition to a working C build environment, you will need automake and libtool on Unices, and cmake on Windows.

Unless you are using a vendored version of hwloc of Windows, you will also need to install pkg-config or one of its clones (pkgconf, pkgconfiglite...), as it is used to find libhwloc and set up hwlocality to link against it.

By default, compatibility with all hwloc 2.x versions is aimed for, which means features from newer versions in the 2.x series (or, in the near future, compatibility with breaking changes from the 3.x series) are not supported by default.

You can enable them, at the cost of losing compatibility with older hwloc 2.x releases, by enabling the cargo feature that matches the lowest hwloc release you need to be compatible with. See the [features] section of this crate's Cargo.toml for more information.

Usage

First, add hwlocality as a dependency:

cargo add hwlocality

Then, inside of your code, set up a Topology. This is the main entry point to the hwloc library, through which you can access almost every operation that hwloc allows.

Here is a quick usage example which walks though the detected hardware topology and prints out a short description of every CPU and cache object known to hwloc:

use hwlocality::{object::depth::NormalDepth, Topology};

fn main() -> eyre::Result<()> {
    let topology = Topology::new()?;

    for depth in NormalDepth::iter_range(NormalDepth::MIN, topology.depth()) {
        println!("*** Objects at depth {depth}");

        for (idx, object) in topology.objects_at_depth(depth).enumerate() {
            println!("{idx}: {object}");
        }
    }

    Ok(())
}

One possible output is:

*** Objects at depth 0
0: Machine
*** Objects at depth 1
0: Package
*** Objects at depth 2
0: L3 (16MB)
*** Objects at depth 3
0: L2 (512KB)
1: L2 (512KB)
2: L2 (512KB)
3: L2 (512KB)
4: L2 (512KB)
5: L2 (512KB)
*** Objects at depth 4
0: L1d (32KB)
1: L1d (32KB)
2: L1d (32KB)
3: L1d (32KB)
4: L1d (32KB)
5: L1d (32KB)
*** Objects at depth 5
0: Core
1: Core
2: Core
3: Core
4: Core
5: Core
*** Objects at depth 6
0: PU
1: PU
2: PU
3: PU
4: PU
5: PU
6: PU
7: PU
8: PU
9: PU
10: PU
11: PU

More examples are available in the source repository.

hwloc API coverage

Most of the features from the hwloc 2.x series are now exposed by hwlocality. But some specialized features, mostly related to interoperability with other APIs, could not make it for various reasons. Issues with the "api coverage" label track unimplemented features, and are a great place to look for potential contributions to this library if you have time!

If you are already familiar with the hwloc C API, you will also be happy to know that #[doc(alias)] attributes are extensively used so that you can search the documentation for hwloc API entities like hwloc_bitmap_t, hwloc_set_cpubind or hwloc_obj::arity and be redirected to the suggested replacement in the Rust API.

The main exceptions to this rule are notions that are not needed in Rust due to ergonomics improvements permitted by the Rust type system. For example...

  • C-style manual destructors are replaced by Drop impls
  • Type argument clarification flags like HWLOC_MEMBIND_BYNODESET are replaced by generics that do the right thing automatically.

License

This project uses the MIT license, please see the LICENSE file for more information.

Dependencies

~0.5–11MB
~112K SLoC