chili

chili

Rust port of Spice, a low-overhead parallelization library

Very low-overhead parallelization primitive, almost identical to rayon::join. At any fork point during computation, it may run the two passed closures in parallel.

It works best in cases where there are many small computations and where it is expensive to estimate how many are left on the current branch in order to stop trying to share work across threads.

Example

The following example sums up all nodes in a binary tree in parallel.

fn sum(node: &Node, scope: &mut Scope) -> u64 {
    let (left, right) = scope.join(
        |s| node.left.as_deref().map(|n| sum(n, s)).unwrap_or_default(),
        |s| node.right.as_deref().map(|n| sum(n, s)).unwrap_or_default(),
    );

    node.val + left + right
}

This is the ideal example since per-node computation is very cheap and the nodes don't keep track of how many descendants are left.

Benchmarks

The following benchmarks measure the time it takes to sum up all the values in a balanced binary tree with varying number of nodes.

AMD Ryzen 7 4800HS (8 cores)

While the improvement over the baseline in the 134M nodes case is close to the theoretical maximum, it's worth noting that the actual time per node is 0.8ns vs. a theoretical 1.8 / 8 = 0.2ns, if we're to compare against the 1K nodes case.

Apple M1 (8 cores)

chili overhead on AMD Ryzen 7 4800HS (8 cores)

The overhead in the 1K nodes case remains approximately constant with respect to the number of threads.