Skip to content

c-cube/moonpool

Repository files navigation

Moonpool

build

A pool within a bigger pool (ie the ocean). Here, we're talking about pools of Thread.t which live within a fixed pool of Domain.t.

This fixed pool of domains is shared between all the pools in moonpool. The rationale is that we should not have more domains than cores, so it's easier to pre-allocate exactly that many domains, and run more flexible thread pools on top.

In addition, some concurrency and parallelism primitives are provided:

  • Moonpool.Fut provides futures/promises that execute on these thread pools. The futures are thread safe.

  • Moonpool.Chan provides simple cooperative and thread-safe channels to use within pool-bound tasks. They're essentially re-usable futures.

    On OCaml 5 (meaning there's actual domains and effects, not just threads), a Fut.await primitive is provided. It's simpler and more powerful than the monadic combinators.

  • Moonpool_forkjoin, in the library moonpool.forkjoin provides the fork-join parallelism primitives to use within tasks running in the pool.

On OCaml 4.xx, there is only one domain; all threads run on it, but the pool abstraction is still useful to provide preemptive concurrency.

Usage

The user can create several thread pools (implementing the interface Runner.t). These pools use regular posix threads, but the threads are spread across multiple domains (on OCaml 5), which enables parallelism.

Current we provide these pool implementations:

  • Fifo_pool is a thread pool that uses a blocking queue to schedule tasks, which means they're picked in the same order they've been scheduled ("fifo"). This pool is simple and will behave fine for coarse-granularity concurrency, but will slow down under heavy contention.
  • Ws_pool is a work-stealing pool, where each thread has its own local queue in addition to a global queue of tasks. This is efficient for workloads with many short tasks that spawn other tasks, but the order in which tasks are run is less predictable. This is useful when throughput is the important thing to optimize.

The function Runner.run_async pool task schedules task() to run on one of the workers of pool, as soon as one is available. No result is returned by run_async.

# #require "threads";;
# let pool = Moonpool.Fifo_pool.create ~num_threads:4 ();;
val pool : Moonpool.Runner.t = <abstr>

# begin
   Moonpool.Runner.run_async pool
    (fun () ->
        Thread.delay 0.1;
        print_endline "running from the pool");
   print_endline "running from the caller";
   Thread.delay 0.3; (* wait for task to run before returning *)
  end ;;
running from the caller
running from the pool
- : unit = ()

To wait until the task is done, you can use Runner.run_wait_block1 instead:

# begin
   Moonpool.Runner.run_wait_block pool
    (fun () ->
        Thread.delay 0.1;
        print_endline "running from the pool");
   print_endline "running from the caller (after waiting)";
  end ;;
running from the pool
running from the caller (after waiting)
- : unit = ()

The function Fut.spawn ~on f schedules f () on the pool on, and immediately returns a future which will eventually hold the result (or an exception).

The function Fut.peek will return the current value, or None if the future is still not completed. The functions Fut.wait_block and Fut.wait_block_exn will block the current thread and wait for the future to complete. There are some deadlock risks associated with careless use of these, so be sure to consult the documentation of the Fut module.

# let fut = Moonpool.Fut.spawn ~on:pool
    (fun () ->
       Thread.delay 0.5;
       1+1);;
val fut : int Moonpool.Fut.t = <abstr>

# Moonpool.Fut.peek fut;
- : int Moonpool.Fut.or_error option = None

# Moonpool.Fut.wait_block_exn fut;;
- : int = 2

Some combinators on futures are also provided, e.g. to wait for all futures in an array to complete:

# let rec fib x =
    if x <= 1 then 1 else fib (x-1) + fib (x-2);;
val fib : int -> int = <fun>

# List.init 10 fib;;
- : int list = [1; 1; 2; 3; 5; 8; 13; 21; 34; 55]

# let fibs = Array.init 35 (fun n -> Moonpool.Fut.spawn ~on:pool (fun () -> fib n));;
val fibs : int Moonpool.Fut.t array =
  [|<abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>;
    <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>;
    <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>;
    <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>; <abstr>;
    <abstr>; <abstr>; <abstr>|]

# Moonpool.Fut.join_array fibs |> Moonpool.Fut.wait_block;;
- : int array Moonpool.Fut.or_error =
Ok
 [|1; 1; 2; 3; 5; 8; 13; 21; 34; 55; 89; 144; 233; 377; 610; 987; 1597; 2584;
   4181; 6765; 10946; 17711; 28657; 46368; 75025; 121393; 196418; 317811;
   514229; 832040; 1346269; 2178309; 3524578; 5702887; 9227465|]

Support for await

On OCaml 5, effect handlers can be used to implement Fut.await : 'a Fut.t -> 'a.

The expression Fut.await some_fut, when run from inside some thread pool, suspends its caller task; the suspended task is then parked, and will be resumed when the future is completed. The pool worker that was executing this expression, in the mean time, moves on to another task. This means that await is free of the deadlock risks associated with Fut.wait_block.

In the following example, we bypass the need for Fut.join_array by simply using regular array functions along with Fut.await.

# let main_fut =
    let open Moonpool.Fut in
    spawn ~on:pool @@ fun () ->
    (* array of sub-futures *)
    let tasks: _ Moonpool.Fut.t array = Array.init 100 (fun i ->
       spawn ~on:pool (fun () ->
           Thread.delay 0.01;
           i+1))
    in
    Array.fold_left (fun n fut -> n + await fut) 0 tasks
  ;;
val main_fut : int Moonpool.Fut.t = <abstr>

# let expected_sum = Array.init 100 (fun i->i+1) |> Array.fold_left (+) 0;;
val expected_sum : int = 5050

# assert (expected_sum = Moonpool.Fut.wait_block_exn main_fut);;
- : unit = ()

Errors

We have a Exn_bt.t type that comes in handy in many places. It bundles together an exception and the backtrace associated with the place the exception was caught.

Fibers

On OCaml 5, Moonpool comes with a library moonpool.fib (module Moonpool_fib) which provides lightweight fibers that can run on any Moonpool runner. These fibers are a sort of lightweight thread, dispatched on the runner's background thread(s). Fibers rely on effects to implement Fiber.await, suspending themselves until the await-ed fiber is done.

# #require "moonpool.fib";;
...

# (* convenient alias *)
  module F = Moonpool_fib;;
module F = Moonpool_fib
# F.main (fun _runner ->
    let f1 = F.spawn (fun () -> fib 10) in
    let f2 = F.spawn (fun () -> fib 15) in
    F.await f1 + F.await f2);;
- : int = 1076

Fibers form a tree, where a fiber calling Fiber.spawn to start a sub-fiber is the sub-fiber's parent. When a parent fails, all its children are cancelled (forced to fail). This is a simple form of Structured Concurrency.

Like a future, a fiber eventually resolves into a value (or an Exn_bt.t) that it's possible to await. With Fiber.res : 'a Fiber.t -> 'a Fut.t it's possible to access that result as a regular future, too. However, this resolution is only done after all the children of the fiber have resolved — the lifetime of fibers forms a well-nested tree in that sense.

When a fiber is suspended because it awaits another fiber (or future), the scheduler's thread on which it was running becomes available again and can go on process another task. When the fiber resumes, it will automatically be re-scheduled on the same runner it started on. This means fibers on pool P1 can await fibers from pool P2 and still be resumed on P1.

In addition to all that, fibers provide fiber local storage (like thread-local storage, but per fiber). This storage is inherited in spawn (as a shallow copy only — it's advisable to only put persistent data in storage to avoid confusing aliasing). The storage is convenient for carrying around context for cross-cutting concerns such as logging or tracing (e.g. a log tag for the current user or request ID, or a tracing scope).

Fork-join

On OCaml 5, again using effect handlers, the sublibrary moonpool.forkjoin provides a module Moonpool_forkjoin implements the fork-join model. It must run on a pool (using Runner.run_async or inside a future via Fut.spawn).

It is generally better to use the work-stealing pool for workloads that rely on fork-join for better performance, because fork-join will tend to spawn lots of shorter tasks.

Here is an simple example of a parallel sort. It uses selection sort for small slices, like this:

# let rec select_sort arr i len =
    if len >= 2 then ( 
      let idx = ref i in
      for j = i+1 to i+len-1 do
        if arr.(j) < arr.(!idx) then idx := j
      done;
      let tmp = arr.(!idx) in
      arr.(!idx) <- arr.(i);
      arr.(i) <- tmp;
      select_sort arr (i+1) (len-1)
    );;
val select_sort : 'a array -> int -> int -> unit = <fun>

And a parallel quicksort for larger slices:

# let rec quicksort arr i len : unit =
    if len <= 10 then select_sort arr i len
    else (
      let pivot = arr.(i + (len / 2)) in
      let low = ref (i - 1) in
      let high = ref (i + len) in

      (* partition the array slice *)
      while !low < !high do
        incr low;
        decr high;
        while arr.(!low) < pivot do
          incr low
        done;
        while arr.(!high) > pivot do
          decr high
        done;
        if !low < !high then (
          let tmp = arr.(!low) in
          arr.(!low) <- arr.(!high);
          arr.(!high) <- tmp
        )
      done;

      (* sort lower half and upper half in parallel *)
      Moonpool_forkjoin.both_ignore
        (fun () -> quicksort arr i (!low - i))
        (fun () -> quicksort arr !low (len - (!low - i)))
    );;
val quicksort : 'a array -> int -> int -> unit = <fun>


# let arr = [| 4;2;1;5;1;10;3 |];;
val arr : int array = [|4; 2; 1; 5; 1; 10; 3|]
# Moonpool.Fut.spawn
    ~on:pool (fun () -> quicksort arr 0 (Array.length arr))
    |> Moonpool.Fut.wait_block_exn;;
- : unit = ()
# arr;;
- : int array = [|1; 1; 2; 3; 4; 5; 10|]


# let arr =
    let rand = Random.State.make [| 42 |] in
    Array.init 40 (fun _-> Random.State.int rand 300);;
val arr : int array =
  [|64; 220; 247; 196; 51; 186; 22; 106; 58; 58; 11; 161; 243; 111; 74; 109;
    49; 135; 59; 192; 132; 38; 19; 44; 126; 147; 182; 83; 95; 231; 204; 121;
    142; 255; 72; 85; 95; 93; 73; 202|]
# Moonpool.Fut.spawn ~on:pool
    (fun () -> quicksort arr 0 (Array.length arr))
  |> Moonpool.Fut.wait_block_exn
  ;;
- : unit = ()
# arr;;
- : int array =
[|11; 19; 22; 38; 44; 49; 51; 58; 58; 59; 64; 72; 73; 74; 83; 85; 93; 95; 95;
  106; 109; 111; 121; 126; 132; 135; 142; 147; 161; 182; 186; 192; 196; 202;
  204; 220; 231; 243; 247; 255|]

Note that the sort had to be started in a task (via Moonpool.Fut.spawn) so that fork-join would run on the thread pool. This is necessary even for the initial iteration because fork-join relies on OCaml 5's effects, meaning that the computation needs to run inside an effect handler provided by the thread pool.

More intuition

To quote gasche:

You are assuming that, if pool P1 has 5000 tasks, and pool P2 has 10 other tasks, then these 10 tasks will get to run faster than if we just added them at the end of pool P1. This sounds like a “fairness” assumption: separate pools will get comparable shares of domain compute ressources, or at least no pool will be delayed too much from running their first tasks.

[…]

  • each pool uses a fixed number of threads, all running simultaneously; if there are more tasks sent to the pool, they are delayed and will only get one of the pool threads when previous tasks have finished
  • separate pools run their separate threads simultaneously, so they compete for compute resources on their domain using OCaml’s systhreads scheduler – which does provide fairness in practice
  • as a result, running in a new pool enables quicker completion than adding to an existing pool (as we will be scheduled right away instead of waiting for previous tasks in our pool to free some threads)
  • the ratio of compute resources that each pool gets should be roughly proportional to its number of worker threads

OCaml versions

This works for OCaml >= 4.08.

  • On OCaml 4.xx, there are no domains, so this is just a library for regular thread pools with not actual parallelism (except for threads that call C code that releases the runtime lock, that is). C calls that do release the runtime lock (e.g. to call Z3, hash a file, etc.) will still run in parallel.

  • on OCaml 5.xx, there is a fixed pool of domains (using the recommended domain count). These domains do not do much by themselves, but we schedule new threads on them, and form pools of threads that contain threads from each domain. Each domain might thus have multiple threads that belong to distinct pools (and several threads from the same pool, too — this is useful for threads blocking on IO); Each pool will have threads running on distinct domains, which enables parallelism.

    A useful analogy is that each domain is a bit like a CPU core, and Thread.t is a logical thread running on a core. Multiple threads have to share a single core and do not run in parallel on it2. We can therefore build pools that spread their worker threads on multiple cores to enable parallelism within each pool.

TODO: actually use https://github.com/haesbaert/ocaml-processor to pin domains to cores, possibly optionally using select in dune.

License

MIT license.

Install

$ opam install moonpool

Footnotes

  1. beware of deadlock! See documentation for more details.

  2. ignoring hyperthreading for the sake of the analogy.