Hi, Hi
There are some bugs with splice in 1.5-dev19... they have been fixed. See this thread for the patches: http://comments.gmane.org/gmane.comp.web.haproxy/12774 (Or google for: "Oh and by the way, the bug was present since 1.5-dev12." ) This is not what Annika is seeing; that bug is about 100% cpu load in userspace haproxy, but Annika is seeing higher system load. Yes system load is much higher then userspace load. Also please tell: - hardware (cpu/ram/nic at least) on old/new cluster - Two Intel(R) Xeon(R) CPU X6550 @ 2.00GHz in each cluster node - 2x Emulex Corporation OneConnect 10Gb NIC (rev 02) in each cluster node - 32gbit RAM in each cluster node - Two nodes per cluster (active-active in the new one) The hardware of the old and the new cluster is the same? Yes. - Debian Squeeze / 3.1.0-1-amd64 / Tickrate 250 - CentOS release 6.4 (Final) / 3.11.5-1.el6 / Tickrate 1000 The higher the tickrate, the higher the CPU load. You quadripled the tickrate, and your load what - quadripled? I suggest you try a lower tickrate in the very same configuration. That said, splice should be way more efficient in 3.11 than in 3.1. Are you using splice-auto or forcing splice by configuring splice-request / splice-response? - We are forcing by splice-request / splice-responce I believe splice is not always more efficient than recv/send; use splice-auto to use it less aggressively (doc: splice-auto): For testing we disabled splicing on one of the cluster members on the new cluster (after succesfull tests). Now load drops below 8 from 16. So I maybe try it with splice-auto and if that does not help with a new haproxy build with the following git commits: http://haproxy.1wt.eu/git?p=haproxy.git;a=commit;h=61d39a0e2a047df78f7f3bfcf5584090913cdc65 http://haproxy.1wt.eu/git?p=haproxy.git;a=commit;h=fa8e2bc68c583a227ebc78bab5779b84065b28da Haproxy uses heuristics to estimate if kernel splicing might improve performance or not. Both directions are handled independently. Note that the heuristics used are not much aggressive in order to limit excessive use of splicing. Don't know if those heuristics are still fully valid for post 3.5 kernels, but it probably doesn't hurt. Regards, Lukas Thank you very much, Annika