ForkJoin

A work stealing fork-join parallelism library.

Inspired by the blog post Data Parallelism in Rust and implemented as part of a master's thesis. Repository hosted at github.com/faern/forkjoin

Library documentation hosted here

This library has been developed to accommodate the needs of three types of algorithms that all fit very well for fork-join parallelism.

Reduce style

Reduce style is where the algorithm receive an argument, recursively compute a value from this argument and return one answer. Examples of this style include recursively finding the n:th Fibonacci number and summing of tree structures. Characteristics of this style is that the algorithm does not need to mutate its argument and the resulting value is only available after every subtask has been fully computed.

In reduce style algorithms the return values of each subtask is passed to a special join function that is executed when all subtasks have completed. To this join function an extra argument can be sent directly from the task if the algorithm has has ReduceStyle::Arg. This can be seen in the examples here.

Example of reduce style (ReduceStyle::NoArg)

use forkjoin::{FJData,TaskResult,ForkPool,AlgoStyle,ReduceStyle,Algorithm};

fn fib_30_with_4_threads() {
    let forkpool = ForkPool::with_threads(4);
    let fibpool = forkpool.init_algorithm(Algorithm {
        fun: fib_task,
        style: AlgoStyle::Reduce(ReduceStyle::NoArg(fib_join)),
    });

    let job = fibpool.schedule(30);
    let result: usize = job.recv().unwrap();
    assert_eq!(1346269, result);
}

fn fib_task(n: usize, _: FJData) -> TaskResult<usize, usize> {
    if n < 2 {
        TaskResult::Done(1)
    } else {
        TaskResult::Fork(vec![n-1,n-2], None)
    }
}

fn fib_join(values: &[usize]) -> usize {
    values.iter().fold(0, |acc, &v| acc + v)
}

Example of reduce style (ReduceStyle::Arg)

use forkjoin::{FJData,TaskResult,ForkPool,AlgoStyle,ReduceStyle,Algorithm};

struct Tree {
    value: usize,
    children: Vec<Tree>,
}

fn sum_tree(t: &Tree) -> usize {
    let forkpool = ForkPool::new();
    let sumpool = forkpool.init_algorithm(Algorithm {
        fun: sum_tree_task,
        style: AlgoStyle::Reduce(ReduceStyle::Arg(sum_tree_join)),
    });
    let job = sumpool.schedule(t);
    job.recv().unwrap()
}

fn sum_tree_task(t: &Tree, fj: FJData) -> TaskResult<&Tree, usize> {
    if t.children.is_empty() {
        TaskResult::Done(t.value)
    } else if fj.depth > fj.workers { // Bad example of serial threshold
        TaskResult::Done(sum_tree_seq(t))
    } else {
        let mut fork_args: Vec<&Tree> = vec![];
        for c in t.children.iter() {
            fork_args.push(c);
        }
        TaskResult::Fork(fork_args, Some(t.value)) // Pass current nodes value to join
    }
}

fn sum_tree_seq(t: &Tree) -> usize {
    t.value + t.children.iter().fold(0, |acc, t2| acc + sum_tree_seq(t2))
}

fn sum_tree_join(value: &usize, values: &[usize]) -> usize {
    *value + values.iter().fold(0, |acc, &v| acc + v)
}

Search style

Search style return results continuously and can sometimes start without any argument, or start with some initial state. The algorithm produce one or multiple output values during the execution, possibly aborting anywhere in the middle. Algorithms where leafs in the problem tree represent a complete solution to the problem (unless the leaf represent a dead end that is not a solution and does not spawn any subtasks), for example nqueens and sudoku solvers, have this style. Characteristics of the search style is that they can produce multiple results and can abort before all tasks in the tree have been computed.

Example of search style

use forkjoin::{FJData,ForkPool,TaskResult,AlgoStyle,Algorithm};

type Queen = usize;
type Board = Vec<Queen>;
type Solutions = Vec<Board>;

fn search_nqueens() {
    let n: usize = 8;
    let empty = vec![];

    let forkpool = ForkPool::with_threads(4);
    let queenpool = forkpool.init_algorithm(Algorithm {
        fun: nqueens_task,
        style: AlgoStyle::Search,
    });

    let job = queenpool.schedule((empty, n));

    let mut solutions: Vec<Board> = vec![];
    loop {
        match job.recv() {
            Err(..) => break, // Job has completed
            Ok(board) => solutions.push(board),
        };
    }
    let num_solutions = solutions.len();
    println!("Found {} solutions to nqueens({}x{})", num_solutions, n, n);
}

fn nqueens_task((q, n): (Board, usize), fj: FJData) -> TaskResult<(Board,usize), Board> {
    if q.len() == n {
        TaskResult::Done(q)
    } else {
        let mut fork_args: Vec<(Board, usize)> = vec![];
        for i in 0..n {
            let mut q2 = q.clone();
            q2.push(i);

            if ok(&q2[..]) {
                fork_args.push((q2, n));
            }
        }
        TaskResult::Fork(fork_args, None)
    }
}

fn ok(q: &[usize]) -> bool {
    for (x1, &y1) in q.iter().enumerate() {
        for (x2, &y2) in q.iter().enumerate() {
            if x2 > x1 {
                let xd = x2-x1;
                if y1 == y2 || y1 == y2 + xd || (y2 >= xd && y1 == y2 - xd) {
                    return false;
                }
            }
        }
    }
    true
}

In-place mutation style

NOTE: This style works in the current lib version, but it requires very ugly unsafe code!

In-place mutation style receive a mutable argument, recursively modifies this value and the result is the argument itself. Sorting algorithms that sort their input arrays are cases of this style. Characteristics of this style is that they mutate their input argument instead of producing any output.

Examples of this will come when they can be nicely implemented.

Tasks

The small units that are executed and can choose to fork or to return a value is the TaskFun. A TaskFun can NEVER block, because that would block the kernel thread it's being executed on. Instead it should decide if it's done calculating or need to fork. This decision is taken in the return value to indicate to the user that a TaskFun need to return before anything can happen.

A TaskFun return a TaskResult. It can be TaskResult::Done(value) if it's done calculating. It can be TaskResult::Fork(args) if it needs to fork.

TODO

• Make mutation style algorithms work without giving join function
• Implement a sorting algorithm. Quicksort?
• Remove need to return None on fork with NoArg
• Make it possible to use algorithms with different Arg & Ret on same pool.
• Make ForkJoin work in stable Rust.
• Remove mutex around channel in search style.