add chunking in poll_write?

[rkuhn]

Since there is poll_flush for triggering a send syscall (which I consider to be expensive), how about buffering data in poll_write up to the chunking threshold?

As an example, the protocol I’m currently working on alternates between writing 4 bytes and some kB. While debugging some issues I saw that the 4 bytes were sent over the wire with a 12 byte header. In my send loop I deliberately call poll_write for as long as I have data, followed by poll_flush and then waiting for more inputs. With some chunking this would probably increase bandwidth by a large factor.

I know from testing that mplex requires flushing, but since yamux currently does not (which I don’t think is documented) we’d probably need to review all protocols for proper use of poll_flush after such a change.

[mxinden]

Off the top of my head, this makes sense to me. While we have optimized this implementation for large payloads (see e.g. #100), we have not done so for small messages.