Conversation

Tony "Abolish ICE" Arcieri

Replying to

@alilleybrinker

Cross-platform async APIs in my opinion have one of the highest bars for running code and precise benchmarks next to a field like cryptography. They are notoriously hard to get right with tradeoffs everywhere

Quote Tweet

Tony "Abolish ICE" Arcieri

The kernel APIs are brittle and unforgiving and to the extent people can actually build cross-platform abstractions it's through blood, sweat, and tears. Nearly every attempt I've seen to second-guess decisions made by people in-the-trenches has ended in a breakage.

Show this thread

1

4

Replying to

and

@alilleybrinker

AIO for storage is also a lot different than the best way to deal with sockets. Making a portable abstraction for any of this involves many sacrifices, and I think at this stage it's hard to end up with a result that's not disappointing for Rust's niche.

Quote Tweet

Replying to @DanielMicay @raggi and @thinkingfish

I don't think it's possible to have an AIO library that's both good and portable right now. The whole point of Rust is having efficient abstractions mapping well to the low-level details. If it's not a thin abstraction over io_uring, it fails at being a good AIO library on Linux.

2

4

Tony "Abolish ICE" Arcieri

Replying to

and

@alilleybrinker

Yeah, there is nothing resembling a portable model for asynchronous filesystem I/O. The best you can do involves throwing thread pools at certain common filesystem operations on certain platforms.

1

3

Replying to

and

@alilleybrinker

nginx uses thread pools for AIO on Linux and it works very well: nginx.org/en/docs/http/n Cloudflare uses a patch for async open in addition to the standard read, write and sendfile support. It's only an issue when you have a huge number of files since otherwise it's cached.

1

3

Replying to

and

@alilleybrinker

There's ongoing work on using io_uring. The main advantage is saving CPU and memory resources. It's not significantly faster for general usage. It's only really significantly faster for lots of small files where the overhead of working with the threads is significantly higher.

1

3

Tony "Abolish ICE" Arcieri

Replying to

and

@alilleybrinker

I'm familiar with io_uring 😜

Quote Tweet

Tony "Abolish ICE" Arcieri

My experience report for shipping a Ruby binding which uses a C extension that wraps io_uring: multiple issue reports from people whose kernel headers were out-of-date

Show this thread

1

3

Replying to

and

@alilleybrinker

Yeah, I saw that. I just meant that in the context of the most popular web server, threads are largely good enough already. I would have thought that io_uring would have more to offer but scaling up the number of threads generally ends up offering comparable performance for it.

1

2

Replying to

and

@alilleybrinker

I'm sure it's a lot more compelling for something like a database. Thread pools work a lot better than people seem to think though. One caveat is nginx allows you to configure separate thread pools for different storage devices via the usual server/location block configuration.

1

2

Replying to

and

@alilleybrinker

I see a lot of problems with Rust's library ecosystem for AIO, etc. but the language portion of it seems really good. I don't really see the problem. I'm glad Rust went this way and I think the approach in Go is going to look like a clear mistake to most people in hindsight.

1

1

4

Replying to

and

@alilleybrinker

If you haven't seen it, look at fuchsia.dev/fuchsia-src/co. This is pretty much the same for Android, where Rust is being heavily adopted for platform code. I doubt they would use Go for anything beyond more tooling if at all. Also github.com/google/kati/bl is a similar experience.

kati/INTERNALS.md at master · google/kati

An experimental GNU make clone. Contribute to google/kati development by creating an account on GitHub.

2

2

Replying to

and

@alilleybrinker

Rust could have easily ended up with GC and M:N threading if it had been rushed along more instead of spending so much time figuring things out. Most of the things I regard as mistakes in Rust are in the standard library, not the language, so they're largely avoidable anyways.

2:14 AM · Mar 11, 2021Twitter Web App

Tony "Abolish ICE" Arcieri

Replying to

and

@alilleybrinker

Shipping the libnative/libgreen split and “libio1” would’ve been bad (not to mention GC)

3