Awesome

Domainslib - Nested-parallel programming

Domainslib provides support for nested-parallel programming. Domainslib provides async/await mechanism for spawning parallel tasks and awaiting their results. On top of this mechanism, domainslib provides parallel iteration functions. At its core, domainslib has an efficient implementation of work-stealing queue in order to efficiently share tasks with other domains.

Here is a sequential program that computes nth Fibonacci number using recursion:

(* fib.ml *)
let n = try int_of_string Sys.argv.(1) with _ -> 1

let rec fib n = if n < 2 then 1 else fib (n - 1) + fib (n - 2)

let main () =
  let r = fib n in
  Printf.printf "fib(%d) = %d\n%!" n r

let _ = main ()

We can parallelise this program using Domainslib:

(* fib_par.ml *)
let num_domains = try int_of_string Sys.argv.(1) with _ -> 1
let n = try int_of_string Sys.argv.(2) with _ -> 1

(* Sequential Fibonacci *)
let rec fib n = 
  if n < 2 then 1 else fib (n - 1) + fib (n - 2)

module T = Domainslib.Task

let rec fib_par pool n =
  if n > 20 then begin
    let a = T.async pool (fun _ -> fib_par pool (n-1)) in
    let b = T.async pool (fun _ -> fib_par pool (n-2)) in
    T.await pool a + T.await pool b
  end else 
    (* Call sequential Fibonacci if the available work is small *)
    fib n

let main () =
  let pool = T.setup_pool ~num_domains:(num_domains - 1) () in
  let res = T.run pool (fun _ -> fib_par pool n) in
  T.teardown_pool pool;
  Printf.printf "fib(%d) = %d\n" n res

let _ = main ()

The parallel program scales nicely compared to the sequential version. The results presented below were obtained on a 2.3 GHz Quad-Core Intel Core i7 MacBook Pro with 4 cores and 8 hardware threads.

$ hyperfine './fib.exe 42' './fib_par.exe 2 42' \
            './fib_par.exe 4 42' './fib_par.exe 8 42'
Benchmark 1: ./fib.exe 42
  Time (mean ± sd):     1.217 s ±  0.018 s    [User: 1.203 s, System: 0.004 s]
  Range (min … max):    1.202 s …  1.261 s    10 runs

Benchmark 2: ./fib_par.exe 2 42
  Time (mean ± sd):    628.2 ms ±   2.9 ms    [User: 1243.1 ms, System: 4.9 ms]
  Range (min … max):   625.7 ms … 634.5 ms    10 runs

Benchmark 3: ./fib_par.exe 4 42
  Time (mean ± sd):    337.6 ms ±  23.4 ms    [User: 1321.8 ms, System: 8.4 ms]
  Range (min … max):   318.5 ms … 377.6 ms    10 runs

Benchmark 4: ./fib_par.exe 8 42
  Time (mean ± sd):    250.0 ms ±   9.4 ms    [User: 1877.1 ms, System: 12.6 ms]
  Range (min … max):   242.5 ms … 277.3 ms    11 runs

Summary
  './fib_par2.exe 8 42' ran
    1.35 ± 0.11 times faster than './fib_par.exe 4 42'
    2.51 ± 0.10 times faster than './fib_par.exe 2 42'
    4.87 ± 0.20 times faster than './fib.exe 42'

More example programs are available here.

Installation

You can install this library using OPAM.

$ opam switch create 5.1.0
$ opam install domainslib

Development

If you are interested in hacking on the implementation, then opam pin this repository:

$ opam switch create 5.0.0+trunk --repo=default,alpha=git+https://github.com/kit-ty-kate/opam-alpha-repository.git
$ git clone https://github.com/ocaml-multicore/domainslib
$ cd domainslib
$ opam pin add domainslib file://`pwd`