Is it problematic to call std::thread::sleep() while iterating a ParallelIterator?

[dan-da]

We have some code that looks kinda like:

 rayon::iter::repeat(0).map_init(|| { ..  }
            |arg, _x| {
                // do some stuff.
                std::thread::sleep(Duration::from_millis(100));
            },
        )
        .find_any(|r| !condition)
        .unwrap();

We also have other code elsewhere that uses rayon iterators (without sleeps).

I'm not sure how the work-stealing works, and I'm wondering if this sleep() can potentially clog up the rayon threadpool and cause the app to effectively hang, with low CPU utlization. That is the behavior we are seeing... app totally stalled, including all tokio tasks, but CPU is basically idle.

So is this a no-no? I haven't seen warnings not to use sleep() in rayon workers, but maybe there should be some?

or its fine?

[cuviper]

std::thread::sleep will block all activity on that thread, including rayon work-stealing.

If you want to give other work a chance to run, consider calling rayon::yield_now() and then adjust your sleep period accordingly. The specifics depend on your reason for sleeping, I suppose.

[dan-da]

that's very helpful. I will give yield_now!() a try. thanks!

[dan-da]

If you want to give other work a chance to run, consider calling rayon::yield_now() and then adjust your sleep period accordingly. The specifics depend on your reason for sleeping, I suppose.

hmm, it's unclear to me how to "adjust your sleep period accordingly" without calling sleep(). yield_now() doesn't accept any duration.

In our case, we are calling sleep() because we have a CPU intensive loop that we want to make CPU friendly when running in dev environment. So we have a cli flag to enable this mode, where the loop runs, but only 10 iterations per second. So our goal would be to enable that functionality using the yield_now!() method.

[dan-da]

I guess one approach would be to calc endtime = now + 100 ms, then call yield_now() in a loop until now >= endtime.

Is there a better way? I fear that could be a very busy loop if no other pending work is executed.

[cuviper]

Looping yield without any sleep would be a busy wait, probably not what you want. You could potentially sleep for some smaller period when yield didn't do anything, until it adds up to your endtime.

It might be better for your task to sleep on some separate dedicated thread. Then you can inject into the rayon pool 10/sec, or just run on that separate thread and let the OS deal with scheduling.

[dan-da]

thx, I think the simplest solution here may be to not use rayon at all for the mode where we want to sleep. after all it is intended for parallelizing work, not waiting around.