unifex task/timed_single_thread_context cannot run 1 million concurrent tasks

[ccotter]

https://github.com/Xavorim/cpp_tasks_bench?tab=readme-ov-file#results-on-my-system benchmarks libunifex against a few other concurrency libraries. The benchmark attempts to launch 1 million concurrent tasks. Unfortunately, the benchmark states libunifex hangs at just 100K tasks. I am able to replicate this myself, with the benchmark program at 100% for a minute until I killed the program.

I reimplemented the benchmark with async_manual_reset_event instead of timed_single_thread_context, and I am able to launch 1 million tasks in about 4 seconds: Xavorim/cpp_tasks_bench#1

I'm entering this issue to see if the timer scheduler can be improved. I'll also compare against stdexec (as the benchmark repo doesn't include it currently).

stdexec benchmark

stdexec appears to be able to run 1 million tasks with exec::timed_thread_context

#include <stdexec/execution.hpp>
#include <exec/timed_thread_scheduler.hpp>
#include <exec/task.hpp>
#include <exec/async_scope.hpp>

#include <chrono>
#include <cstdio>
#include <iostream>
#include <cstdlib>
#include <ranges>

exec::timed_thread_context TIMER;

auto async_sleep(auto s) { return exec::schedule_after(TIMER.get_scheduler(), s); }

int main(int argc, char **argv) {
    if (argc < 2) {
        fprintf(stderr,
                "Please specify the number of tasks\n"
                "example: %s 10000\n",
                argv[0]);
        return EXIT_FAILURE;
    }

    exec::async_scope scope;
    for (auto i : std::views::iota(0, std::stoi(argv[1]))) {
        scope.spawn([]() -> exec::task<void> { co_await async_sleep(std::chrono::seconds(10)); }());
    }

    stdexec::sync_wait(scope.on_empty());
    return 0;
}

$ time ./a.out 1000000

real    0m17.098s
user    0m11.241s
sys     0m4.150s

[ericniebler]

I'll also compare against stdexec (as the benchmark repo doesn't include it currently).

stdexec doesn't have an async_manual_reset_event yet, but it probably wouldn't be hard to port the one from libunifex.

[ccotter]

Ah ok, libunifex's timed_single_thread_context enqueues into a sorted linked list, and enqueueing each item must traverse each element to find the correct location to store the item. stdexec uses an intrusive heap, so enqueueing 1 million items is far more efficient.

[ericniebler]

I wonder how important the single thread context is. Who uses it and for what?

[ispeters]

timed_single_thread_context is the default scheduler for Meta users who need a TimeScheduler so it gets production use, but I don't think we're loading it with a million items under any normal circumstances.