server: N-threads event-based (epoll) just accepting our pre-made payload and nothing more with cpu-affinity on each of these threads.
client: minimal http 1.1 complient chunk (totally static w/ in-memory fd sending with splice - see http://man7.org/linux/man-pages/man2/vmsplice.2.html). We can probably epoll on these writes and use the same idea as in the server
What are we looking for?
how much we lose on middleware processing (haproxy)
affects that this can make to the lb in terms of QoS
On the receiving sides (both parties have a sending and a receiving side) we can read the packets but drop entirely the content by using MSG_TRUNC (with recv instead of just read)
server: N-threads event-based (epoll) just accepting our pre-made payload and nothing more with cpu-affinity on each of these threads. client: minimal http 1.1 complient chunk (totally static w/ in-memory fd sending with splice - see http://man7.org/linux/man-pages/man2/vmsplice.2.html). We can probably epoll on these writes and use the same idea as in the server
What are we looking for?