Oneshot perfromance

[tkaitchuck]

I have a rather large and somewhat complex application that is getting completely bottlenecked on the performance of oneshots.

High level overview of what's happening:

1. oneshot is created and returned to caller.
2. Onshot is passed through an mpsc to a previously spawned task.
3. unrelated stuff happens. (involves another handoff with an mpsc)
4. Same task from step 2 completes the oneshot.

The caller calls the function to create the oneshot and then awaits it and then awaits the oneshot over and over in a loop.

Flame graph is attached. (unfortunately git doesn't allow svgs so it's just a PNG zoomed in)

What's odd about that is that the same code path actually uses two mpscs and those barely register on the flamegraph at all. I tried just replacing the oneshots with mpscs but actually was slower.

Is there some way to speed up this sort of "creating, using, and dropping" oneshots in a loop? Maybe a way to avoid the wake_by_ref call which for whatever reason don't seem to be happening with mpscs when they are used for many messages?

[Darksonn]

Did you try out other channel implementations? Some options are:

• futures crate
• async-channel crate
• flume crate

[tkaitchuck]

I tried the futures crate. The perf was similar. I'll look into the other two.