[pcwalton]

> Over the years we've so internalized the fact that threads are expensive that we forgot it's an implementation detail that could actually be fixed.

This is something I wish more people would realize. We assume that 1:1 threads inherently can't scale, mostly because of folklore. (As far as I can tell, the origin of this claim is back-of-the-envelope calculations involving stack sizes, leading to address space exhaustion on 32-bit—obviously something that hasn't been relevant for servers for a decade.) That's why we build async/await and elaborate userspace scheduling runtimes like the one Go has (and the one Rust used to have pre-release).

I think it's worth questioning the fundamental assumptions that are causing us to do this. Can we identify why exactly 1:1 threading is uncompetitive with async/await schemes, and fix that?

All this, by the way, is not to say that Rust made the wrong decision in focusing on async/await. Rust generally follows the philosophy of "if the computing environment is difficult, it's Rust that has to adapt".

[pron]

> Can we identify why exactly 1:1 threading is uncompetitive with async/await schemes, and fix that?

I think we can identify it, but fixing it is not easy, at least at the kernel level.

The easy part is the cost of scheduling. The kernel must schedule threads with very different behaviors -- say, encoding a video or serving requests over a socket -- so it uses a scheduling algorithm that's a compromise for all uses. But we can let the language's runtime take the scheduling of kernel threads with something like this: https://youtu.be/KXuZi9aeGTw

The harder part is managing the memory required for the stack. Even on 64-bit systems, the OS can't shrink and grow stacks finely enough. For example, I don't think there's a hard requirement that, say, memory below the sp is never accessed, so the OS can't even be sure about how much of the stack is used and can't uncommit pages. But even if there were such a requirement, or, that the language could tell the OS it follows such a requirement, still the OS can only manage memory at a page granularity, which is too much for lightweight concurrency. Any finer than that requires knowing about all pointers into the stack.

We can do it in languages that track the location of all pointers in all frames, though, which is what we're attempting to do in Java, and this allows us to move stacks around, grow them and shrink them as necessary, even at a word granularity.

> All this, by the way, is not to say that Rust made the wrong decision in focusing on async/await. Rust generally follows the philosophy of "if the computing environment is difficult, it's Rust that has to adapt".

... and it follows C++'s (horribly named) "zero-cost abstractions" philosophy which can reasonably be said to be a requirement of Rust's target domains. So I certainly don't think async/await is wrong for C++/Rust/Zig, but I think it's wrong for, say, JavaScript and C#, and we're going a different way in Java (more like Scheme's).

Another possible contributing factor is that in one respect Rust is a higher-level language than Java or JS: it compiles to a VM -- LLVM/WASM -- over which it doesn't have full control, so it doesn't have complete control over its backend. That, BTW, is why Kotlin adopted something similar to async/await.

[pcwalton]

> at a page granularity, which is too much for lightweight concurrency

I don't think that's been conclusively shown. 4kB is smaller than a lot of single stack frames. It would be interesting for someone to measure how large e.g. Go stacks are in practice—not during microbenchmarks.

> Another possible contributing factor is that perhaps ironically, in one respect Rust is a higher-level language than Java or JS, as it compiles to a VM -- LLVM/WASM -- over which it doesn't have full control, so it doesn't have complete control over its backend.

It's not really a question of "control"; we can and do land changes upstream in LLVM (though, admittedly, they sometimes get stuck in review black holes, like the noalias stuff did). The issue is more that LLVM is very large, monolithic, and hard to change. Upstream global ISel is years overdue, for example. That's one of the reasons we have Cranelift: it is much smaller and more flexible.

But in any case, the code generator isn't the main issue here. If GC metadata were a major priority, it could be done in Cranelift or with Azul's LLVM GC support. The bigger issue is that being able to relocate pointers into the stack may not even be possible. Certainly it seems incompatible with unsafe code, and even without unsafe code it may not be feasible due to pointer-to-integer casts and so forth. Never say never, but relocatable stacks in Rust seems very hard.

The upshot is that making threading in Rust competitive in performance to async I/O for heavy workloads would have involved a tremendous amount of work in the Linux kernel, LLVM, and language design—all for an uncertain payoff. It might have turned out that even after all that work, async/await was still faster. After all, even if you make stack growth fast, it's hard to compete with a system that has no stack growth at all! Ultimately, the choice was clear.

[pron]

> I don't think that's been conclusively shown.

That would be interesting to study. We'll do it for Java pretty soon, I guess.

> After all, even if you make stack growth fast, it's hard to compete with a system that has no stack growth at all!

If the same Rust workload were to run on such a system there wouldn't be stack growth, either. The stack would stay at whatever size Rust now uses to store the async fn's state.

> a tremendous amount of work in the Linux kernel

I don't think any work in the kernel would have been required.

> Ultimately, the choice was clear.

I agree the choice is pretty clear for the "zero-cost abstractions" approach, and that Rust should follow it given its target domains. But for domains where the "zero-cost use" approach makes more sense, the increase in accidental complexity is probably not worth some gain in worst-case latency. But that's always the tradeoff the two approaches make.

[pcwalton]

You suggested the switchto patch, right? That requires Linux kernel work.

[pron]

Assuming you know how to relocate stacks, you could implement efficient delimited continuations in Rust without the kernel's involvement, and use whatever it is you use for async/await scheduling today (so I'm not talking about some built-in runtime scheduler like Go's or Erlang's). But this would require changes in LLVM, and also a somewhat increased footprint (although the footprint could be addressed separately, also by changes in LLVM). The kernel work is only required if you want to suspend computations with non-Rust frames, which you can't do today, either.

[pcwalton]

That sounds like mioco/old M:N. That was slower than our current 1:1 threading.

[littlestymaar]

> So I certainly don't think async/await is wrong for C++/Rust/Zig, but I think it's wrong for, say, JavaScript and C#, and we're going a different way in Java (more like Scheme's).

How can it be wrong for JavaScript? Because JavaScript is a single threaded environment, it was the only possible choice and it's one of the most influential change in the language (in itself it changed more the language use than the whole ES6 bundle). It's a really great success and I'm not sure why you'd like to revert it.

[pron]

> I'm not sure why you'd like to revert it.

I don't. The old way and async/await aren't the only two options.

> it was the only possible choice

Why?

[littlestymaar]

What would you do instead? Use a green thread system on top of a single-threaded VM? Great, now you have two kinds of blocking calls: the one which only block its own thread, and the one which block all threads at the same time because it blocks the VM itself. How ergonomic!

Remember, the single threaded character of the js VM is not an implementation detail, it's part of the spec. Hate it or love it but the web works this way.

Also, if you think threads are a good concurrency abstraction, let's play a little game: consider you need to read 1M files on a spinning disk. How many threads do you need to run on to get the maximum read performance:

a) one per file

b) just one

c) one per core

d) a magic number which depends on your hard disk's firmware and your workload.

Convenient and intuitive right?

Threads are a dated concurrency primitive which would have died long ago if wasn't also a good parallelism primitive.

[pron]

> Great, now you have two kinds of blocking calls: the one which only block its own thread, and the one which block all threads at the same time because it blocks the VM itself. How ergonomic!

You could only have the first kind, but it's the same situation with async/await. Only async/await tracks the "good" kind of blocking, yet lets the "bad" kind go untracked.

> How many threads do you need to run on yo get the maximum read performance

The exact same number as you would for doing it with `await Promise.all` The same knowledge you have about the scheduling mechanism doesn't go away if you're no longer required to annotate functions with `async`.

> Threads are a dated concurrency primitive which would have died long ago if wasn't also a good parallelism primitive.

Maybe they are, but async/await are the exact same construct only that you have to annotate every blocking function with "async" and every blocking call with "await". If you had a language with threads but no async/await that had that requirement you would not have been able to tell the difference between it and one that has async/await.

[littlestymaar]

> Only async/await tracks the "good" kind of blocking, yet lets the "bad" kind go untracked.

The “bad kind” is indistinguishable from CPU intensive computation anyway (which cannot be tracked), but at least you have a guarantee when you are using the good kind. (Unfortunately, in JavaScript, promises are run as soon as you spawn them, so they can still contain a CPU heavy task that will block your event loop, Rust made the right call by not doing anything until the future is polled).

> The exact same number as you would for doing it with `await Promise.all`

From a user's perspective, when I'm using promises, I have no idea how it's run behind (at it can be nonblocking all the way down if you are using a kernel that supports nonblocking file IOs). This example was specifically about OS threads though, not about green ones (but it will still be less expensive to spawn 1M futures than 1M stackful coroutines).

> Maybe they are, but async/await are the exact same construct only that you have to annotate every blocking function with "async" and every blocking call with "await". If you had a language with threads but no async/await that had that requirement you would not have been able to tell the difference between it and one that has async/awaitof

I don't really understand your point. Async/await is syntax sugar on top of futures/promises, which itself is a concurrency tool on top of nonblocking syscalls. Of course you could add the same sugar on top of OS threads (this is even a classic exercise for people learning how the Future system works in Rust), that wouldn't make much sense to use such thing in practice though.

The question is whether the (green) threading model is a better abstraction on top of nonblocking syscalls than async/await is. For JavaScript the answer is obviously no, because all you have behind is a single threaded VM, so you lose the only winning point of green threading: the ability to use the same paradigm for concurrency and parallelism. In all other regards (performance, complexity from the user's perspective, from an implementation perspective, etc.) async/await is just a better option.

[kps]

> something that hasn't been relevant for servers for a decade

Has Rust officially relinquished small systems to C?

[pcwalton]

I explicitly said that async I/O is the right decision for Rust. And C has no I/O concurrency features at all, so that's completely irrelevant.