In the interest of sparking a little discussion (not too spicy please, we’re having a nice clean start here) I thought I would ask the question. It’s something I’m legitimately wondering about as someone who has reached for tokio by default for years.
I’m aware of async-std and smol, probably unaware of others. If you’ve used or prefer a different async runtime, what trade-offs might I be interested in or what features am I missing out on?
You must log in or register to comment.
Not so much relevant for production use cases but from an educational point of view: whorl is pretty great. One file that explains async runtime basics from top to bottom. The cleverness of doing this in one source code file is remarkable.
There are executors for more specific use cases.
- For example,
bastionis a “Highly-available Distributed Fault-tolerant Runtime”, inspired by Erlang, and including an async executor. embassyincludes an async executor specifically for embedded systems.fuchsia-asyncis an executor for use in the Fuchsia OS.wasm-bindgen-futuresconverts RustFutures to JavaScriptPromises and schedules them to run to completion. It could also be seen as a (very basic) executor.
I think gtk-rs is another interesting example. The
glibcrate provides an async executor, since glib already requires running its main loop. So if you just need to run a couple futures you don’t need another executor in one or more other threads, and tasks spawned in the glib executor can call functions that need to be run in that thread, since gtk types are generally not thread safe.So integrating into other things could be a reason to use a different executor…
- For example,
This won’t really help you consider another runtime specifically, but it is an instance of not choosing Tokio.
When I was benchmarking with Tokio I found it had quite a difference (higher cpu usage) to other options I was exploring. My specific use-case was an async single-threaded application which does mostly network io + timers. I would be interested to see how Tokio fairs today (I tested it over a year ago) and quite possibly any advice or tips in writing efficient code with it. My benchmark was send/recv with a UDP socket and print out stats in a timer (that’s the core idea of what I needed to use it with). Comparing both CPU usage and maximum throughput.
For normal async, I would probably still reach for Tokio, but the use case I was looking at was very much specific to what I was working on at my job (still am), and lower CPU usage as well as higher throughput is what we needed. We ended up building a small internal runtime that uses mio, it was also fun to implement. In the benchmarks we ran, it used less CPU % than Tokio, and achieved higher throughput, by enough that we thought it was worth continuing to use our internal one.
