Graeme, Turned out it was a mixture of conntrack on the real servers filling up,
http://www.austintek.com/LVS/LVS-HOWTO/HOWTO/LVS-HOWTO.performance.html#conntrack_filling_tables and fairly old hardware on the load balancer (I didn't realise just how old).... Thanks. 2009/3/12 Graeme Fowler <[email protected]>: > On Thu, 2009-03-12 at 09:18 +0000, Malcolm Turnbull wrote: >> When the active connections climb to 60,000 and throughput is about >> 100MB+ we get a sudden and massive drop of in throughput... > > That looks to me very much like an interrupt-handling issue. What > architecture are your NICs using - PCI, PCIe etc etc? Are they on board > using a common bus? > > I'd surmise that either they're on a common PCI bus which is saturated, > or they're sharing an IRQ and just can't go any faster. Can you separate > them (either in BIOS or driver)? > > You might want to look back at CPU affinity which was discussed > recently, although you say this a Celery so that throws out SMP magic as > a fix! > > Graeme > > > _______________________________________________ > Please read the documentation before posting - it's available at: > http://www.linuxvirtualserver.org/ > > LinuxVirtualServer.org mailing list - [email protected] > Send requests to [email protected] > or go to http://lists.graemef.net/mailman/listinfo/lvs-users > -- Regards, Malcolm Turnbull. Loadbalancer.org Ltd. Phone: +44 (0)870 443 8779 http://www.loadbalancer.org/ _______________________________________________ Please read the documentation before posting - it's available at: http://www.linuxvirtualserver.org/ LinuxVirtualServer.org mailing list - [email protected] Send requests to [email protected] or go to http://lists.graemef.net/mailman/listinfo/lvs-users
