Asynchronous Programming in Rust book
rust-lang.github.io
rust-lang.github.io
I’ll show myself out.
Desync takes a slightly different approach to asynchronous programming: instead of being based around the idea of scheduling operations on threads, and then synchronising data across those threads, it's based on the idea of scheduling operations on data.
There's only two basic operations: 'desync' runs an operation on some data in the background, and 'sync' runs one synchronously.
All operations are run in order and 'sync' returns a value so it's a way to retrieve data from an asynchronous operation. It's sort of like setting up some threads with some data protected by a mutex and sending results between them using mpsc channels, except without the need to build any of the scaffolding. ('sync' also makes borrowing data from one task to use in another effortless)
https://rust-lang.github.io/async-book/getting_started/state...
Is missing, somewhat ironic?
Feels very much like the state of async matches the state of the guide. :P
What is the state of async? Is it close? Is it still changing with the futures 0.3-beta not finalized?
Are we six months away? A year?
What is going on with futures 0.3? Why is everyone still using 0.1?
How does that relate to these issues?
It superficially appears like the whole async story is still in a concept stage...
Futures 0.3 uses nightly only-features that are landing in Rust within (hopefully) the next two releases of Rust. Namely, Futures 0.3 is a way to experiment with async programming using the async/await syntax.
> Why is everyone still using 0.1?
Futures 0.3 is still in flux—but settling down in recent weeks—and is relying on nightly-only features. Futures 0.1 is used heavily in Hyper and Tokio, but we intend to move to Futures 0.3/std::future::Future when they're available on stable or shortly thereafter.
(The Tokio and Hyper projects take backwards compatibility _extremely_ seriously.) (Disclaimer: I help maintain Tokio/Hyper, but am nowhere near are prolific as the main authors.)
AFAIK becaue tokio isn't upgrading until some issues are ironed out. As long as tokio is on 0.1 the rest of the community will be on 0.1 too.
Commercial products tend to avoid this. Sales of version N go way down before version N+1 is generating revenue. Overall revenue drops during the transition. That's not good.
(Also, there's a compatibility layer, so even the people that want to play with the shiny new N + 1 can do so, even though it's not explicitly directly supported.)
How does async translate calls to other async functions? Is refactoring into smaller async functions less efficient? If not, how does it deal with (possibly indirect) recursive function calls? Does it give up or select a loop breaker?
And what is the purpose of the pingpong between executor->Waker->push onto executor?
I am also still unsure what the approach to multithreading might be. Multiple executors with work stealing or one dispatch executor with worker threads or something else still?
I believe it builds the calls up into one larger future, so it shouldn't be any less efficient.
I can't answer any of the other questions with any certainty.
There's nothing special going on. Remember, async on a function is something like
async fn function(argument: &str) -> usize {
fn function(argument: &str) -> impl Future<Item=usize> {
> If not, how does it deal with (possibly indirect) recursive function calls?
Recursive calls to async functions will fail to compile: https://github.com/rust-lang/rust/issues/53690
That said, see that discussion; the trait object form will probably eventually work.
Heavy recursion isn't generally Rust's style, since we don't have guaranteed TCO, so you threaten to overflow the stack and panic.
> Is refactoring into smaller async functions less efficient?
That's a complicated question. It really depends. I don't think it should be, thanks to inlining, but am not 100% sure.
> And what is the purpose of the pingpong between executor->Waker->push onto executor?
Right now, the best resource is https://boats.gitlab.io/blog/post/wakers-i/ and https://boats.gitlab.io/blog/post/wakers-ii/
> I am also still unsure what the approach to multithreading might be.
You have options! Tokio now does multiple executors with work-stealing by default, in my understanding.
From the second blog post I actually found https://github.com/tokio-rs/tokio/pull/660 which switched tokio from 1 reactor+worker threads to n reactors with work stealing.
Skimmed through https://vorner.github.io/async-bench.html. If I understand it correctly, one get about twice the performance with async.
Is this correct? Seems like a compromise (code complexity vs performance) not worth taking.
If so, I'd argue that long term once async/await have landed properly, the code largely looks and behaves the same. With that said, I've not even used it yet, because I've got no clue when this is landing enough that I can reasonably use it.. and I'm on Nightly lol.
Yes, the code the developer needs to read, write and understand.
I'm not familiar of how async/await will be in Rust, but I guess some code differences/complexities can be:
1. Make sure, manually(?), that all things are async / non-blocking.
2. Implementing Future.poll / wrapping types in Future? (What is Pin? ref https://rust-lang.github.io/async-book/execution/future.html)
3. Async polution, a function that uses async must be async too?
4. Setup some scheduler that maintain how many concurrent async operations one thread has?
5. More verbose error-messages / stack-traces?
Coming from JS, that's a non-problem in Rust. You can easily make a function blocking by creating a event loop and resolving the future you get from another function in it. So when I refactor my code to be async, I'm starting by making a single function async, and the moving the event loop from function to function, until as much of the code is async as I want.
In hindsight, I prefer async/await. My reason is primarily that like your example points out, it really lets me be in full control over the scheduled behavior. I could even take non-io work and make it "async". Ie, some long processing application takes a break every million iterations to let other tasks steal some work. That's just cool!
Arguably a similar thing could be designed in Go if every million iterations you used some type of IO primitive, like sending some data over a channel, but the behavior of Rust's model is more fine grained.
> 1. Make sure, manually(?), that all things are async / non-blocking.
You'd have to make sure any IO you do is using Futures - ie, use a package to provide async IO primitives for disk and network access. You would also need to use the appropriate await syntax call on any future using methods - that would require a bit of overhead to know, but at least the compiler has your back on that.
> 2. Implementing Future.poll / wrapping types in Future?
In most cases I don't think you'd have a use case to implement a Future - would you? Ie, main IO calls are the big ones for wasting threads - and libraries like mio/hyper/etc provide your IO primitives.
> 3. Async polution, a function that uses async must be async too?
Yea, my understanding is that this is definitely an issue. I am already planning on using `async` tags on basically all my functions, because everything I use bound to IO in one form or another.
On the bright side, I believe (don't quote me!) that you can drop ugly `fn foo() -> Futures<Item=Result<A,B>>` wrapping, since I believe `async fn foo() -> Result<A,B>` does the same thing. .. again, the syntax is not finalized haha.
> 4. Setup some scheduler that maintain how many concurrent async operations one thread has?
If you're using Async I'd imagine you'd already have chosen a scheduler. I believe Tokio will be the defacto - though Rayon might be involved here, not sure.
> 5. More verbose error-messages / stack-traces?
Errors themselves would be unaffected, if you're talking normal error values - remember those are just values in Rust, like Go, so not much special there.
Though as you said, I imagine if you dump a trace it would look different, no idea.
None of this post was meant to counter you in anyway. I just hoped to provide a bit of clarity on the tiny things I can contribute to. I hope I helped more than hurt. Have a nice day :)
And, for web servers, it can be more than 2x. For example, look at techempower's plaintext benchmark: https://www.techempower.com/benchmarks/#section=data-r17&hw=...
Hyper gets 7,013,819. It's async. Iron gets 109,815, and is synchronous. That's 63x. Iron uses hyper under the hood, so that should be a good comparison.
But you are correct, if you don't have a specific need, async is generally harder than using threads for concurrency. Ideally the async/await work in Rust is going to make that trade-off less extreme than it is today, which may mean more people will feel comfortable using it as it should reduce boiler plate.
Could you expand on that? I've never heard that mentioned about async before.
(I also misinterpreted the context as I read the top level comment as async vs a single thread with blocking code, but after rereading it that makes more sense.)
Yes, a thread pool, consisting of one thread per core/computing unit. The units of work are then scheduled between the threads. Units of work here being some kind of IO, e.g. servicing HTTP requests.
> but if you want to spin up a few hundred thousand of them...
Hm. Thought there was a limit for work that can be done concurrently by the CPU, based on the number of cores/hyper-threads available. Found this on threads and IO performance [1], it seems to make the same point.
What kind of work load is common to spread over so many threads (on the same machine)? Does the OS switch efficiently between hundred of threads on regular CPUs? Genuinely interested.
1: https://www.jstorimer.com/blogs/workingwithcode/7970125-how-...
> Thought there was a limit for work that can be done concurrently by the CPU,
Right. But in an IO bound scenario, the CPU isn't doing work; it's waiting on IO. So, because threads are generally heavy, you don't want a ton of them, taking up memory, doing nothing.
But, when you have lightweight threads, you can spin up one per connection. This ends up being simpler, and you don't have the large memory usage. This is what nginx does, in a sense. It still has a worker per core, but each of those workers can handle thousands of requests simultaneously, because it's all non-blocking.
That limit to concurrent work is exactly why non-blocking architectures are so important, and task systems fit into them really nicely.
Now you do this 1000 times you use barely any resources but the server uses 1GB in stack allocations which might be the maximum and noone else can connect.
Tokio is a good, default choice, but some projects may have different needs.
Tokio itself is great though, so if you have no strong reason not to use it, I'd recommend it.