Improve

[dinosaure]

We just pre-allocated buffers to read/write and we improve a bit the code. We are also be able to catch-up earlier when the user wants to stop the server.

[hannesm]

I've a hard time figuring out the allocation behaviour before and after, so now we preallocate? how big are the buffers being allocated? are those for each connection read and write direction?

[dinosaure]

We can split the improvement into 3 parts:

1. the buffer size, here we enlarge the buffer from 2047 bytes to 16383 bytes, you can see the diff here.
2. we pre-allocate a read buffer and a write buffer to copy between bigarray and bytes, you can see the diff here.
3. and we set the socket with TCP_NODELAY, you can see the diff here

From my experiment, it's a win but I'm currently try to provide a reproducible benchmark between httpcats, vif and httpun+eio.

[dinosaure]

I just need to keep in my mind that the second point (use the same buffer with Bytes.unsafe_to_string. see here does not work (probably) for mhttp (but we can figure out about an extra allocation with the underlying layer Flow).