OoO En ce doux début de matinée du samedi 15 octobre 2011, vers 08:05, Willy Tarreau <[email protected]> disait :
>> > Still, 40 ms of latency. I have also modified stud to not send the >> > proxy line until there is something else to send (by commenting out >> > "ev_io_start" in "if (OPTIONS.WRITE_PROXY_LINE)" section of >> > handle_connect(). No change. Since the code handling PROXY protocol is >> > so small in stud, I have no other idea of what to do. See here for the >> > relevant code: >> > https://github.com/bumptech/stud/blob/master/stud.c#L489 >> >> > Any idea where the latency could lie? >> >> The problem was stud did not disable Nagle's algorithm on the backend >> side. Therefore, nothing to do with HAProxy. > Interesting, thank for the info. This means we all need to re-run tests > then :-) Did you add a setsockopt(TCP_NODELAY) or was there just an option > for that ? This was solved by Jamie in this commit: https://github.com/bumptech/stud/commit/7ad212a3f8a1416b8ac997a54e2dd58840f75851 -- Vincent Bernat ☯ http://vincent.bernat.im Make input easy to prepare and output self-explanatory. - The Elements of Programming Style (Kernighan & Plauger)
