Ralf Weber wrote: > Moin!
gack! > ... > > So as said we agree that the Community should put more effort in > getting the Linux people to do better UDP stack instead of switching > DNS to TCP, which other people have been suggesting. on a 1.4-GHz opteron in 2003 or so, recvmsg/sendmsg was giving us ~30K UDP TPS. there was no recvmmsg() at that time. this was freebsd 5 or so. so we wrote a user mode UDP implementation that was based on BPF, knowing that the syscall interface used by libpcap received up to 50 packets per call. we got 200K UDP TPS. so, i'd be very curious to see a linux-vs-freebsd UDP bakeoff (just using recvmsg/sendmsg) since i know that both kernels have had a great deal of performance work done to their UDP stacks in the past decade. and i'd be very curious to see a recvmsg-vs-recvmmsg UDP bakeoff, since if there isn't a huge speedup from using recvmmsg, then the locking in the kernel is still too course-grained. (this test would have to be run separately on linux and on freebsd, since they obviously use different kernels.) -- Paul Vixie _______________________________________________ dns-operations mailing list [email protected] https://lists.dns-oarc.net/mailman/listinfo/dns-operations dns-jobs mailing list https://lists.dns-oarc.net/mailman/listinfo/dns-jobs
