#smart-pointers #parallel #data-structures

spinout

A thread-safe clonable and mutable smart-pointer akin to Arc<Mutex<T>>, using spin-lock instead of system futex

3 releases

0.0.12 Sep 28, 2022
0.0.11 Sep 28, 2022
0.0.1 Sep 28, 2022

#724 in Concurrency

48 downloads per month

MIT/Apache

20KB
352 lines

Atom

This is an experimental implementation, meaning API might change and bugs might arise.

Experimental API for a synchronous and mutable smart-pointer type Atom<T> and its Weak<T> variant. Also includes the associated spin-lock type SpinLock.

Unlike Mutex<T>, Atom<T> does not use system futexes, but instead uses a simple spin-lock. This can be advantageous in cases of low contention i.e. when the lock is only held for a short time and there are few threads competing for the lock.

Usage

Atom<T> mimics the behavior of Arc<Mutex<T>>, but with a slightly different API. Instead of lock-guards we simply access the inner T inside a closure.

let atom = Atom::new(5);
atom.lock(|x| *x += 5);
assert_eq!(atom.get(), 10);

We can access parts of the inner value and map parts of it to a new value:

let atom = Atom::new(vec![1, 2, 3]);
let sum: i32 = atom.map(|x| x.iter().sum());
assert_eq!(sum, 6);

Or mutate and map:

let atom = Atom::new(vec![1, 2, 3]);
let three = atom.map_mut(|x| x.pop());
assert_eq!(three, Some(3));
assert_eq!(atom.get(), vec![1, 2]);

Example

use spinout::Atom;

fn main() {

    let mut numbers = vec![];

    for i in 1..43 {
        numbers.push(i);
    }

    let numbers = Atom::new(numbers);
    let t1_numbers = numbers.clone();
    let t2_numbers = numbers.clone();

    let results = Atom::new(vec![]);
    let t1_results = results.clone();
    let t2_results = results.clone();

    let t1 = std::thread::spawn(move || {
        while let Some(x) = t1_numbers.map_mut(|x| x.pop()) {
            if x % 2 == 0 {
                t1_results.lock(|v| v.push(x));
            }
        }
    });

    let t2 = std::thread::spawn(move || {
        while let Some(x) = t2_numbers.map_mut(|x| x.pop()) {
            if x % 2 == 0 {
                t2_results.lock(|v| v.push(x));
            }
        }
    });

    t1.join().unwrap();
    t2.join().unwrap();

    let mut results = results.get();

    results.sort();

    let expected = [
        2, 4, 6, 8, 10, 12, 14, 16,
        18, 20, 22, 24, 26, 28, 30,
        32, 34, 36, 38, 40, 42
    ];

    assert_eq!(results, expected);
}

Benchmarks

These tests we run on a AMD Ryzen 3 3100 4-Core Processor using the Criterion statistical benchmarking tool.

There are four different tests that simulate a real world scenario with a small thread count and low contention.

In the Balanced Read and Write and Read Heavy Read and Write we do a sort on a small vector and read the biggest value from the (reversed) vector.

Balanced Read and Write Read Heavy Read and Write

In Write Only only test we do a sort and a reverse on a small vector.

Write Only

In the Read Only test we do a find on a value in the vector.

Read Only

No runtime deps