Thanks everyone, looks like i'll have to benchmark myself (which is fine) but I'm always afraid because I know "proper benchmarking is hard. (TM)"

Feel free to throw any other side advice in. I'm looking to get a broad perspective on this.

Straight up shutting off the garbage collector in exchange for memory is an interesting concept. But I wonder how quickly it will get eaten up. ALSO, if I do shut it off, when i turn it on is it going to take much longer? Does the GC take linear / quadratically more time based on the N of "items need to be freed"?

The thing is, I'm looking to parallelize a dedicated server for running MASSIVE numbers of objects (10000's or 100000's) at high tick rate. Worse, for my particular "game mode", the "rounds" can last hours. So while a slow crawl of RAM is okay, it won't be okay if it hits 30 GB in an hour. So that's going to be another "it works IF it works [in our particular game/application]" as opposed to "this definitely will/won't work". That's more "benchmarking and see" scenarios, which I'm trying to avoid as much as possible. You normally don't want to willfully START a project with a design where the only way to know if it works... is to bench.

There's also an option of periodically firing off (without ending the game) say, once every hour and tell everyone to just "live with it" because it's (I HOPE!) not going to be a 15/30/60 second delay. In my particular application, that still wouldn't be that disruptive. (Unless turning GC off hits max ram every couple minutes. Then again nobody is going to be okay with that.)

On Saturday, 4 January 2020 at 11:30:53 UTC, dwdv wrote:
Creates a Task on the GC heap that calls an alias.

If possible, there's also scopedTask, which allocates on the stack: https://dlang.org/phobos/std_parallelism.html#.scopedTask

So my question is: Has anyone done any analysis over how "dangerous" it is to use GC'd tasks for _small_ tasks (in terms of milliseconds)?

Nothing major, but https://github.com/mratsim/weave/tree/master/benchmarks/fibonacci puts quite a bit of pressure on various implementations. You might want to profile ./pfib 40:

import std;

ulong fib(uint n) {
    if (n < 2) return n;

    auto x = scopedTask!fib(n-1); // vs. Task!fib(n-1);
    taskPool.put(x);
    auto y = fib(n-2);
    return x.yieldForce + y; // {yield,spin,work}Force
}

void main(string[] args) {
enforce(args.length == 2, "Usage: fib <n-th fibonacci number requested>");
    auto n = args[1].to!uint;
    // defaultPoolThreads(totalCPUs);
    writefln!"fib(%d) = %d"(n, fib(n));
}

At least D isn't locking up beyond 12 tasks; looking at you, stdlib-nim. :)


Reply via email to