Hi! I have to admit, my previous "details" were somewhat vague - by "tons of traffic" i meant exactly what you described - a huge number of shortlived connections per second (some millions in a few seconds)
On Sun, May 9, 2021 at 5:27 PM Maxim Dounin <mdou...@mdounin.ru> wrote: > Hello! > > On Sat, May 08, 2021 at 10:01:19AM +0200, Charlie Kilo wrote: > > > Thanks Maxim for further explaining! > > > > We do listen to a huge number of IPs with tons of traffic and huge spikes > > on them. We really need to avoid any type > > of congestion, therefore the reuseport. > > > > While many of the ip:port combos are simply there for failover purposes > and > > actually aren't "in use", I see right now > > no feasible way to reduce the number of listening sockets before the > > upgrade and restore them afterwards. That would be hugely complex, > > error-prone and would still leave us with a window where the instant > > failover wouldn't work as expected. > > Thanks for the details. > > Note that "reuseport" doesn't help with "tons of traffic". It can > help in the special case when there are a lot of new connections > are established per second, and all these connections are > short-lived, so a large part of CPU time is spent in establishing > new connections. > > Further, "reuseport" doesn't help if there are already multiple > listening sockets, and new connections are already distributed > between these sockets. Its main use is when you have only one > listening socket and want to reduce lock contention on this socket > by providing additional listening sockets. > > The only case when "reuseport" is unavoidable in nginx now is when > you want to handle UDP proxying with sessions. > > -- > Maxim Dounin > http://mdounin.ru/ > _______________________________________________ > nginx mailing list > nginx@nginx.org > http://mailman.nginx.org/mailman/listinfo/nginx >
_______________________________________________ nginx mailing list nginx@nginx.org http://mailman.nginx.org/mailman/listinfo/nginx
