co_await vs goroutines

[unicomp21]

Anybody know of benchmark data anywhere? I'm wondering if co_await is indeed faster than goroutines.

[emaxerrno]

There is. Talk called negative cost abstractions in CPP con which is amazing. The tldr is yes because the compiler generates a state machine.

[unicomp21]

@senior7515 many thanks. wow, i've always thought it would be nice to have an "await" for cache lines misses. had no idea that's exactly what's going on w/ co_await. amazing.

this makes me want to use the redpanda c++ client even more, lol. from a code generation perspective, i'm tempted to believe some pretty amazing things could be done here.

[BenPope]

I've been meaning to do some benchmarks and throw this into some loops for a while now: https://github.com/scylladb/seastar/blob/master/include/seastar/core/prefetch.hh

[unicomp21]

Starting to feel like a kid standing outside a candy store, lol. @BenPope I'm also curious, do the spectre mitigations, etc. added to the kernel, have an impact on this sort of optimization? Or is that unrelated?

My thinking, in the bare metal, non multi-tenant case, the spectre mitigations (and others) are unnecessary, so it might warrant a special kernel build w/ them all disabled?

[BenPope]

Good question, I haven't looked specifically.

There are some details in this blog, which I haven't read this in a while: https://www.scylladb.com/2018/01/07/cost-of-avoiding-a-meltdown/

TLDR; not as much as many other types of IO intensive software.

[unicomp21]

Random thought. @BenPope does llvm support co_await when compiling to webassembly?

[BenPope]

Also a good question. I don't have an answer. I believe webassembly supports async, but we're getting into the weeds.

There's nothing inherently stopping the coroutine model from being supported, but I don't know the specifics, or the level of support for transpilation of particular languages in the various implementations, or the level of support of coroutines in llvm IR (vs the clang frontend)

Our coprocessors implementation uses V8, to give you a starting point.

Ideally we'd ban loops that aren't async, for the purpose of cooperative multitasking, but I don't know where we are with the implementation.

[unicomp21]

@senior7515 any idea if llvm can compile the co_await state machine to webassembly?

[emaxerrno]

I don’t see why not. Id try it on the recent llvm 13.

Note we still find issues from time to time with the compiler, so mileage may vary