Hacker News Reader

Comment by ThePhysicist | Hacker News Reader

Just got started with Rust and yeah, my biggest pain point so far were lifetimes and async code. I finally replaced tokio with multithreaded blocking code, which seems to be much simpler and more familiar. The problem with async was that some of the libraries I needed (e.g. for QUIC) didn't support it, so you had an unholy mixture of blocking and non-blocking code.

Coming from a C++ background I appreciate Rust's novel approach of borrow checking, though I'm not sure if it's more "elegant" than C++'s approach with unique pointers, move constructors etc.. What I love is the modern ecosystem and build tooling around the language (including documentation and package management), for me that's Rust's main advantage over C++.

View on HN

I've successfully replaced tokio with the lower-level mio in one such case. After all threads and async serve different purposes: Threads when you're computing a lot, async when you're waiting a lot, no?

Threads work OK for concurrency in many apps. The developer cost of async may not be worth it in many cases.

I would say that threads work _very well_ for concurrency in _most_ apps. Thread context switches have gotten much much cheaper over the years, and the idea that threads are heavyweight is very outdated.

True, java apps have comfortably run 10s of thousands of threads without issue historically.

Coroutines are nice as they allow trivially writing an app which scales to millions of concurrent operations. Rust is nice as it lets us write code for every occasion, async is awful as it poisons the entire ecosystem. The same result could have been achieved with a few tactically inserted yield statements for I/O and locking which are pretty much the only times that async or coroutines help.

I'm just an average web developer, so I could be wrong here, but that's not my understanding of it.

Your application has n threads to begin with, if you use async, these will be utilized to schedule tasks. You're basically trusting your language to properly pause and restart the procedures.

If you spawn new threads, these will be managed by your kernel. If that kernel is Linux, that it already has a pretty good algorithm to pause and restart them, so you're not really wasting a lot of resources.

So I believe there are effectively two differences: threads can scale up to the limit of your os/kernel and hardware. Async can scale up to the level of the threads your language spawns to handle these tasks. Threads trust the kernel/OS to prioritize workloads, async trusts the language

The important difference is in purpose, not implementation. Most async runtimes use a thread or more per core, and the OS is doing that thread scheduling, so it's not the either-or situation that you describe.

The interesting thing about async (especially async-await) is that it is a fluent way to write a program that waits for events (e.g. I/O), without resorting to inefficient designs like thread-per-client in order to achieve concurrency.

Software (even in C) has made use of this concept for a long time through poll() and friends; async-await is just a higher level abstraction over the same concept.

> The important difference is in purpose, not implementation.

I couldn't disagree stronger on that one. Abstraction is generally good, but this becomes a purely philosophical discussion as soon as you ignore the actual implementation when talking about the differences of how something actuals behaves and should be used. And philosophical discussions like that have their place but are imo pointless in the context of what should be used, when. They're generally better placed in a context of deciding wherever you want to implement an abstraction

I do fundamentally agree with the rest of your comment however. that's what I meant with trusting your language to prioritize workloads vs your OS.

I'm not saying that implementation doesn't matter. I'm saying that that there isn't a fundamental implementation difference "async" vs "threads". Threads are used in both cases. It's a different choice of abstractions over the same underlying concepts.

FYI, Rust (the language) has no concept of prioritising workloads; there is no async runtime in the language.

Even with any of the async runtimes, workloads (i.e. tasks doing work) are prioritised by the OS. It's just threads executing code. Async runtimes are primarily concerned with the gaps between work, like waiting for I/O or other events.

> If you spawn new threads, these will be managed by your kernel. If that kernel is Linux, that it already has a pretty good algorithm to pause and restart them, so you're not really wasting a lot of resources.

Thread context switches are pretty expensive. If you use an async runtime such as tokio, all your tasks will be spawned on a fixed number of threads and you won't have any context switches on a task switch.

    Thread context switches are pretty expensive.

Opinions are pretty divided on this one these days. There was a time when this was very much true but there is some evidence that it is no longer, practically speaking, an issue except in extreme cases. You do pay a cost but it's not immediately obvious that the cost is material. I reach for a threadpool first and async when it becomes clear that I need the extra performance. 99% of the time the threadpool is more than good enough.

> After all threads and async serve different purposes: Threads when you're computing a lot, async when you're waiting a lot, no?

It doesn't have to be one or the other. Go uses coroutines and multiplexes them onto multiple OS-level threads (usually as many as there are cores).

That's what tokio does under the hood.

Just use `Arc`, it's fine, it will be plenty fast.

That's what I ended up doing, Arc<Mutex<...>> for sharing stuff between threads, works really well.

I have recently written a blog post about `Arc` and `Mutex` if you want to get an in-depth view on using them: https://itsallaboutthebit.com/arc-mutex/

I always wonder about code bases using bolt-on reference counting. I would count C++ shared_ptr and Rust's Arc among them. At this point wouldn't it be better to just use a regular GC language? What I don't like in modern C++ is that almost everything seems to be heap allocated, when it certainly doesn't need to be. In case of GC language you at least get fast allocations and heap compaction. But it is true that resource release is more predictable with RC. On the other hand you get those cascades of releases in RC.

As always it is a game of trade offs, but it is something that bothers me.

> At this point wouldn't it be better to just use a regular GC language?

No, because you wouldn't slap `Arc` on every single variable. The example from the blog post is a dispatcher and `Arc` is then needed for dispatched functions. The same might be true for queues and other stuff where you need to move stuff between threads, but other than that it would be mostly `Arc` free.

As an example: when you write a web service in one of the frameworks like Actix, you typically use `Arc` for stuff that you need to share between all of the handlers like DB connections, queues, counters etc. All of the other stuff would be relatively straightforward Rust without much consideration for lifetimes.

Which QUIC library did you use? Quinn (the one I work on) comes with an async/await interface.

Quiche.

Reply on news.ycombinator.com