Makes perfect sense - but with openvswitch 1.7.1 I'm seeing stuff like: 2013-04-06T20:03:55Z|6338190|timeval|WARN|82 ms poll interval (24 ms user, 52 ms system) is over 52 times the weighted mean interval 2 ms (2423146772 samples) 2013-04-07T03:48:04Z|6378252|timeval|WARN|context switches: 0 voluntary, 3 involuntary 2013-04-08T08:08:30Z|6589218|poll_loop|WARN|wakeup due to [POLLIN] on fd 18 (unknown anon_inode:[eventpoll]) at ../lib/dpif-linux.c:1183 (93% CPU usage)
CPU usage above 90% on 2.4Ghz E5645 with ~300MB/s and ~300K packets/s. (Noticeable SSH latency - sub-second though) CPU usage above 60% with ~120MB/s and 100K packets. I know storage traffic will be jumbo frames and might not be as difficult to handle as normal traffic in OpenVSwitch - but I'd like to minimize the latency for the storage as much as possible. Any hints on what to do here, something looks wrong or is this expected ? Anybody able to recommend running storage with / without openvswitch ? Thanks Regards Kristoffer On 05/04/2013, at 19.00.43, Peter Phaal <peter.ph...@inmon.com> wrote: > I can't speak to the performance impact of running storage traffic over OVS. > We have storage running over OVS in our small XenServer pool and haven't > seen any ill effects, but that isn't much of a test. However, the sFlow > instrumentation in OVS gives useful visibility into storage activity, for > example, looking at AoE traffic: > > http://blog.sflow.com/2011/03/aoe.html > > Regards, > Peter > > On Apr 5, 2013, at 1:32 AM, Kristoffer Egefelt <kristof...@itoc.dk> wrote: > >> Thanks for the input - I may not have explained myself properly though - I'm >> not considering pci-passthrough. >> >> What I would like to confirm is, if there's any problems running storage >> traffic over OVS, latency/performance wise - in this case using SR-IOV VFs >> inside dom0, to seperate the traffic. >> >> What I think I need is two seperate networks on a XEN host - one for >> Openvswitch and one for NFS storage traffic. >> The problem here is that I then need to create 2xLACP bonds on the network >> cards - which may not be possible. >> >> The easy way would be to have just one LACP bond, connect OVS to it, and >> create a OVS interface for the storage and others for the VMs - but I'm >> having doubts if it will impact performance. >> >> Do you guys run NFS / iSCSI / AOE traffic over OVS or on seperate interfaces >> (physical or with SR-IOV(Intel) or nPAR (Broadcom)? >> >> Thanks ;-) >> >> Regards >> Kristoffer >> >> >> On 05/04/2013, at 07.41.22, Ramana Reddy <gtvrre...@gmail.com> wrote: >> >>> You can use 10Gb NIC cards, but when you assign vnic to each VM, the >>> traffic is directly goes through VNIC, and >>> not through OVS. The observed behaviour is that if the traffic goes via >>> OVS, you may not see 10Gb speed, and it varies. What is the need of sending >>> traffic via OVS. Is there any decision making happening at OVS side. >>> >>> >>> >>> On Thu, Apr 4, 2013 at 6:39 PM, Kristoffer Egefelt <kristof...@itoc.dk> >>> wrote: >>> Hi, >>> >>> I want to use 10Gig Intel x520 NICs - should I: >>> >>> - Run storage (iscsi/nfs) over OVS ? >>> - Create VFs and run storage and OVS on separate interfaces ? >>> - Buy more physical NICs even if I don't need the bandwidth ? >>> >>> Any experiences with SR-IOV, storage latency or other issues to suggest one >>> over the other ? >>> >>> Thanks >>> >>> Regards >>> Kristoffer >>> _______________________________________________ >>> discuss mailing list >>> discuss@openvswitch.org >>> http://openvswitch.org/mailman/listinfo/discuss >>> >>> _______________________________________________ >>> discuss mailing list >>> discuss@openvswitch.org >>> http://openvswitch.org/mailman/listinfo/discuss >> >> _______________________________________________ >> discuss mailing list >> discuss@openvswitch.org >> http://openvswitch.org/mailman/listinfo/discuss > > _______________________________________________ > discuss mailing list > discuss@openvswitch.org > http://openvswitch.org/mailman/listinfo/discuss
_______________________________________________ discuss mailing list discuss@openvswitch.org http://openvswitch.org/mailman/listinfo/discuss
