On Nov 5, 2013, at 10:15 AM, Ben Pfaff <[email protected]> wrote: > On Mon, Nov 04, 2013 at 02:44:38PM -0800, Jarno Rajahalme wrote: >> Add a FAQ categorry "Performance Problems". So far the only entry >> addresses the issue with using a new kernel module with an older >> (pre-megaflows) userspace. >> >> Signed-off-by: Jarno Rajahalme <[email protected]> > > What if we just name the section "Performance"? I can see a demand for > broad explanations of OVS performance, as well as descriptions of > performance problems. >
OK. >> +Performance Problems >> +-------------------- >> + >> +Q: I just upgraded and I see a performance drop. Why? >> + >> +A: Sometimes support for new protocol fields is added to OVS. If a >> + new OVS kernel module is paired with an older (pre 1.11) OVS >> + userspace, there may be a preformance drop. In particular, when >> + the new fields concern existing and widely used protocols, old >> + implementations may face a drop in performance. Since OVS release >> + 1.11, which added kernel wildcarding, this should be less of an >> + issue, as any new fields not understood by the userspace remain >> + automatically wildcarded. >> + >> + For example, OVS release 2.1 added support for TCP flags matching. >> + Pairing a new OVS kernel module with OVS userspace without kernel >> + wildcarding, each TCP flags combination for each TCP connection >> + appears as a separate flow miss for the OVS userspace, multiplying >> + the required flow setups in most (TCP-heavy) workloads. >> + >> + Upgrading the OVS userspace components will resolve these >> + performance issues. > > To me, this seems pretty difficult to understand. I understand it, > because I know the underlying issues, but I am not sure that a casual > reader would. > > Perhaps we could rephrase it by starting the answer out with an > explanation that some kernel/userspace combinations do not perform as > well as others, that it is best to pair the same versions of each, and > then list very clearly the combinations that perform worst. Also, it is > probably worth mentioning not just OVS versions but upstream kernel > versions (since a lot of users would prefer to avoid compiling a > separate module if they can). Where can I find out which kernels come with which OVS kernel modules? > An actual explanation of the technical > reasons for the performance differences is nice, but it's probably not > what most users really want when they read this FAQ. OK. Jarno _______________________________________________ dev mailing list [email protected] http://openvswitch.org/mailman/listinfo/dev
