Unify Channel Dependency

[ibraheemdev]

Hyper currently depends on both futures_channel and tokio::sync, using both channel types in different places. To cut down on dependencies one of these should be removed.

[seanmonstar]

hyper currently uses different channels because the workloads of each benefit from the different channels. When I tried to settle on just one, the benchmarks got worse.

[ibraheemdev]

Oh interesting, I didn't see anything documented so I just assumed it was oversight.

I assume there's little difference between the performance of the oneshot channels. Digging deeper, the main use of futures-channel is for a 0 capacity channel, which tokio does not support. I assume you tried to switch everything to tokio and saw performance degradation, which makes more sense because tokio's channels should generally perform better.

[ibraheemdev]

futures_channel's mpsc::channel(0) is equivalent to a channel with the capacity of the number of senders, so tokio can actually be used instead there as well. Would you be okay with everything switching to tokio if there are equivalent benchmark results?

[seanmonstar]

Yes certainly!

[ibraheemdev]

Hm, it seems doing things would require using tokio_util::PollSender, which would add another dependency anyways.

[GlenDC]

Does this need prior work or can it just be implemented and compare the existing benchmarks before and after?

[seanmonstar]

I don't know of any other prior work.

[redactedontop]

I believe it'd be benefitial to use PollSender as if @ibraheemdev is correct, Tokio's channels should be faster.