[Question] Is it intentional that static_thread_pool is basically spinlocking when idle?

[miwechner]

Hi,

Thanks for your work, I am really enoying working with S&R.

I wanted to use exec::static_thread_pool in a GUI application and noticed that I have very high CPU usage when I create an exec::static_thread_pool without enqueuing anything. I think it comes down to this function which means one stealer thread is always kept alive and attempts to pop the queues, even though nothing has been enqueued.

stdexec/include/exec/static_thread_pool.hpp

Lines 930 to 934 in b888185

	inline void static_thread_pool_::thread_state::clear_stealing() {
	if (pool_->numThiefs_.fetch_sub(1, std::memory_order_relaxed) == 1) {
	notify_one_sleeping();
	}
	}

I guess this is intentional to reduce latency, but when i started i have to say that this behaviour wasn't clear to me. I also could not find any documentation for static_thread_pool or issue here on Github talking about this. So i am asking just in case for me and for anyone else that would attempt the same.

Thanks,

Michael

[maikel]

Lets say its at least a known issue. It was discussed in issue #1305 and your question is another motivational push to change that.

[maikel]

Hi Michael,

I made a PR where the pool only wakes up new thiefs if at least one victim is awake. If it is the last thread it will be put to sleep.

Can you try that, whether it fits better?

Thank you for reporting the issue!