Controlling async actions that might conflict in F#

问题

I have many actions (Async<T> list) to perform in F#. I can execute most of these actions in parallel, but some might conflict due to file locks etc.

For each action, I can generate a "key" (int) that determines if the actions might conflict:

If action a has key i and action b has key j and i = j, then a and b might conflict. They must be executed serially.
If action a has key i and action b has key j and i <> j, then a and b will never conflict. They may be executed in parallel.

I would like to execute my actions (int * Async<T>) list in an efficient way and without conflicts.

I imagine that the process would be something like:

Group all actions by key
Chain each group serially into one Async
Run each chain in parallel

How can I implement this in F#?

How are these problems usually handled?

My attempt at a fully sequential implementation:

let wrapTasks<'T> (tasks : (int * Async<'T>) list) : Async<'T list> = async {
  return
    tasks 
    |> Seq.map (fun (k, t) -> t |> Async.RunSynchronously)
    |> Seq.toList
}

回答1:

With a helper function taking a 'promise' for a value x and one for a set of values acc:

module Async =
    let sequence x acc = async {
        let! x = x
        let! y = acc
        return x :: y
    }

we can asynchronously group the tasks by their 'lock id', clean up the resulting list a bit and then sequence each group into a single async that 'contains' the list of the results of its group. This list is then processed in parallel. Once ts : 'b list [] is available, we flatten it:

let wrapTasks tasks = async {
    let! ts =
        tasks
        |> List.groupBy fst
        |> List.map (snd >> List.map snd)
        |> List.map (fun asyncs -> List.foldBack Async.sequence asyncs (async { return [] }))
        |> Async.Parallel
    return ts |> List.ofArray |> List.collect id
}

This can be tested with e.g.

List.init 50 (fun i -> i % 5, async {
    let now = System.DateTime.UtcNow.Ticks
    do! Async.Sleep 10
    return i, now })
|> wrapTasks
|> Async.RunSynchronously
|> List.groupBy snd
|> List.map (fun (t, rs) -> t, rs |> List.map fst)
|> List.sort

By varying the divisor we can adjust the level of parallelism and convince ourselves that the function works as expected :-)

  [(636766393199727614L, [0; 1; 2; 3; 4]);
   (636766393199962986L, [5; 6; 7; 8; 9]);
   (636766393200068008L, [10; 11; 12; 13; 14]);
   (636766393200278385L, [15; 16; 17; 18; 19]);
   (636766393200382690L, [20; 21; 22; 23; 24]);
   (636766393200597692L, [25; 26; 27; 28; 29]);
   (636766393200703235L, [30; 31; 32; 33; 34]);
   (636766393200918241L, [35; 36; 37; 38; 39]);
   (636766393201027938L, [40; 41; 42; 43; 44]);
   (636766393201133307L, [45; 46; 47; 48; 49])]

_{Full disclosure: I had to execute the test a few times for getting this nice result. Usually numbers will be a bit off.}

回答2:

This is a possible solution:

let wrapTasks (tasks : (int * Async<'T>) list) =
    tasks
    |> List.groupBy fst 
    |> Seq.map (fun (k, ts) -> async {
        for (i, t) in ts do
            let! r = t
            ()
    })
    |> Async.Parallel
    |> Async.RunSynchronously

来源：https://stackoverflow.com/questions/53092395/controlling-async-actions-that-might-conflict-in-f

标签

concurrency

async-await