Girls just wanna have fast MPMC queues with bounded waiting | Hacker News Reader

Girls just wanna have fast MPMC queues with bounded waiting

nahla.dev

203 pointsEvgeniyZh46 comments

> Disclaimer: An earlier version of this post claimed the structure is wait-free, this is incorrect. Being wait-free requires that failure or suspension of any thread can’t cause failure or suspension of another thread. This queue in fact does not fulfill that requirement. The main section which discusses the wait bounds of queue operations has been amended to reflect this, but other parts of this article have not been. As such there may parts of the text which refer to this as a wait-free queue, which it is not. I chose to keep those sections to avoid rewriting chunks of this post after it was already posted. Thanks for the correction Reddit user matthieum!

Classy disclaimer! matthieum's (long) reddit comment is also an informative read: https://www.reddit.com/r/rust/comments/1up0uhg/girls_just_wa...

Thanks. I jumped at the headline. I'd be happy with wait-free MPSC. I haven't checked in for a while. Have there been any breakthroughs in low-complexity wait-free queues in the past 10 years?

The closest thing I know of, is that there was a concurrent queue algo called LCRQ

It originally required double-width CAS, but IIRC in recent years someone figured out how to remove this to make it more portable

Best reference I could find from cursory google:

https://ppopp23.sigplan.org/details/PPoPP-2023-papers/2/The-...

https://nikitakoval.org/publications/ppopp23-lprq.pdf seems to be the paper in question.

This paper [0] from 2022 is pretty good. "Low complexity" it is not, though.

[0] https://arxiv.org/pdf/2201.02179

I suspect the search space of low-complexity, or at least what I'd consider "low-complexity", wait-free queues is pretty much exhausted at this point.

The MPSC/MPMC structure in Aeron is wait-free with respect to producers - one producer cannot block another

There are simple node based algorithms that achieves a similar guarantee:

https://web.archive.org/web/20240928080729/https://www.1024c...

There is also a MPMC algorithm on this site very similar to the article

https://web.archive.org/web/20220524214823/https://www.1024c...

Nice to see Vyukov's MPMC queue mentioned. It's pretty neat. I have used a C implementation[1] of this in a small personal project.

[1]: https://github.com/dorjoy03/dsync/blob/master/src/mpmc_queue...

Both Vyukov queues are fast and useful in practice, but neither is even obstruction-free, let alone lock-free or wait-free.

What's important is you know the trade-offs you are making.

You can't have a bounded queue that is always non-blocking because slow consumers can block producers.

You can't have a global FIFO order + multiple producers without slow producers blocking consumers.

You can't have a global FIFO order + have have non-atomic reserve and commit without a interrupted/de-scheduled producer thread being able to block the consumer

If you want atomic commit then you lose separate reserve which means either unbounded memory or atomic fixed-size data with sentinel values, ABA problems etc.

There are trade-offs everywhere, and it's best to pick the data structure that fits your needs just like any other problem.

> There are trade-offs everywhere, and it's best to pick the data structure that fits your needs just like any other problem.

That part I think is most crucial. Neither "Lock-free" nor "Wait-free" are vague terms for how awesome a thing is, they're specific properties which are expensive to provide, if you need such a property it was indispensable, if you don't need it then you can likely do better without it.

Exactly. I mentioned that those queues aren't formally obstruction-free because the context of the conversation was new developments in wait-free queues, even though I have only needed the guarantee once in my career and end up using descendants of the Vyukov MPMC cycle queue in practically all other cases because they are better on the metrics that count, like speed.

What was the one time when you need something wait free? I'm assuming interacting with hardware?

Hard real-time industrial automation. Worst job ever, by the way.

I don't think any of my immediate opportunities are "real-time industrial automation" shaped, but, can't hurt to ask: Do you think it was the work that inherently was the worst ever, or was there some other enviromental factor (e.g. terrible wages, for some reason your office was up six flights of stairs with no elevator, colleague insisted on playing music from your least favourite genre at maximum volume) ?

It was the work itself. Embedded systems are already known for bad tooling. Once safety certification gets thrown in the mix you're limited to ancient broken versions of GCC that have been rebranded and certified, and are thus more fit for purpose than newer versions of GCC that actually work. MISRA, its derivatives, and the associated static analysis tooling ecosystem are also low quality evolutionary dead ends that ignore the state of the art in both academia and industry.

Not a ton, the inexorable march to wider scalar dispatch, deeper pipelines, and ever less uniform geometries (I pine for the halcyon days of two NUMA modes with a QPI bar) has made asymmetrical fencing juicy enough to be be worth the squeeze at the margins, you weren't seeing a ton of `asm volatile ""` on one side and `membarrier()` on the other a decade ago and you'll see that now.

But I think Nathan Bronson's work out of IIRC Standford about 10 or 15 years ago is still more or less the canvas you paint on.

Surprised nobody here nor on the reddit thread caught this one: https://github.com/nahla-nee/wfqueue/blob/main/src/lib.rs#L3...

    unsafe impl<T, const N: usize> Sync for WFQueue<T, N> {}
    unsafe impl<T, const N: usize> Send for WFQueue<T, N> {}

These impls are unsound, because neither constrains `T` to be `Sync`/`Send`. As-written this would let you declare a `WFQueue<Rc<T>, N>` and pass non-atomic-refcount pointers between threads.

The fix is straightforward:

    unsafe impl<T: Sync, const N: usize> Sync for WFQueue<T, N> {}
    unsafe impl<T: Send, const N: usize> Send for WFQueue<T, N> {}

I.e., WFQueue is only Sync if T is Sync, and likewise for Send.

Actually, later on, the code makes a similar mistake, but only for one impl.

    unsafe impl<'a, T, const N: usize> Send for DrivableWFEnqueue<'a, T, N> where T: Send {}
    unsafe impl<'a, T, const N: usize> Send for DrivableWFDequeue<'a, T, N> {}

WFQueue isn't handing out &T, so Send is sufficient:

  unsafe impl<T: Send, const N: usize> Sync for WFQueue<T, N> {}
  unsafe impl<T: Send, const N: usize> Send for WFQueue<T, N> {}

Here's my widely used implementation of this approach in C++: https://github.com/rigtorp/MPMCQueue

This was the very best bounded MPMC queue when I last looked into these things years ago, and as far as descendants of the Vyukov MPMC cycle queue go, I don't think it's possible to do much better.

I think your citation date is off, by the way. As far as I can tell, it was first published in January 2011.

On the topic of lock free data structures I found this one on a SPSC very interesting too https://david.alvarezrosa.com/posts/optimizing-a-lock-free-r... taking it from 12M to 305M ops/s

That looks like a rewrite of my earlier work: https://rigtorp.se/ringbuffer/

https://github.com/andersc/fastqueue2/tree/main

Perhaps I missed it but there didn't appear to be discussion of false sharing between the N individual data slots. It might be beneficial to pad each slot to a cache line width (or at least less slots per line), and/or using some kind of bijective hashing on the slot lookup so that sequential tickets don't access adjacent slots.

You would use one of those approaches:

If you align and pad each slot there won't be any false sharing and the stream prefetcher can kick in if there's only one producer or consumer.

If you use bijective hashing you reduce false sharing without aligning and padding. This can save memory at the expense of the stream prefetcher never kicking in.

>fast MPMC queues with bounded waiting

Sounds like fun.

It's very rare (if ever) I'd want M on both ends... and no work stealing. Usually it's one M and an S.

Reply on news.ycombinator.com