Hacker News Reader

Comment by conradludgate | Hacker News Reader

I don't think I'd recommend using this in production. The benchmarks look good, but by immediately waking the waker[0], you've effectively created a spin-lock. They may work in some very specific circumstances, but they will most likely in practice be more costly to your scheduler (which likely uses locks btw) than just using locks

[0]: https://github.com/fortress-build/whirlwind/blob/0e4ae5a2aba...

View on HN

+1. I'd be curious how much of a pessimization to uncontended workloads it'd be to just use `tokio::sync::RwLock`.

and, if we want to keep it as a spinlock, I'm curious how much the immediate wakeup compares to using `tokio::task::yield_now`: https://docs.rs/tokio/latest/tokio/task/fn.yield_now.html

This is an interesting idea. I am gonna try this out - especially with dashmap, I think that could perform very well.

You could also look into shamelessly "taking inspiration" from async-lock.

I don't believe that waking the waker in `poll` synchronously waits / runs poll again immediately. I think it is more likely just adding the future back to the global queue to be polled. I could be wrong though, I'll look into this more. Thanks for the info!

It does immediately put itself into the queue to be polled again. But that's no different in effect to a spin-lock. If you have other tasks in your runtime, this will be putting excess pressure on your scheduler

Expanding on this. If you have a lot of concurrent tasks, you will overflow[0] the task local queue and be bottlenecked by the global queue mutex[1]

[0]: https://github.com/tokio-rs/tokio/blob/8897885425bf3d89053f8... [1]: https://github.com/tokio-rs/tokio/blob/8897885425bf3d89053f8...

Oh this is really good to know, thank you!

Hmm, imo it's definitely better than directly spinlocking to have many spinlocks running cooperatively, but you're right that it may not be ideal. Thanks for pointing this out. I'll see if I can find a better way to coordinate the polling/waking of lock acquisition futures.

Tokio has a task budget that will cap at 32 or 256 polls of the same task before switching to another task. So, yes a spinlock, but not likely to deadlock.

Yeah, after looking into this more I think this was a big oversight on my part. Working on a (hopefully) better way of doing this right now - I'm thinking per-shard waker queues and only falling back to spinlocking like this if the queues are full.

Another offender is AtomicWaker [1] which does the same on contention.

[1] https://docs.rs/atomic-waker/latest/atomic_waker/

AtomicWaker is much less bad, because it will only spin in the case that another thread has spuriously called `wake` – i.e. called `wake` despite that fact that the future it’s waking is in fact unable to progress.

In the common case, the waker side will operate some logic like:

    set_flag();
    atomic_waker.wake();

and the future will run:

    atomic_waker.register(cx.waker());
    if flag_is_set() {
        Poll::Ready(())
    } else {
        Poll::Pending
    }

Thus, even if `register` decides to “spin”, the flag will already be set, and so the future will not spin in reality (it may be polled unnecessarily, but this will only happen once).

I can’t immediately think of examples where support for spurious waking is necessary.

The producers of an MPSC queue may use an AtomicWaker::wake to signal to the consumer that a new item is ready. In this case all wake-ups are necessary (not spurious).

Reply on news.ycombinator.com