Date: Mon, 17 Aug 2015 09:54:36 -0700 From: Maxim Sobolev <sobomax@FreeBSD.org> To: Luigi Rizzo <rizzo@iet.unipi.it> Cc: Babak Farrokhi <farrokhi@freebsd.org>, "Alexander V. Chernikov" <melifaro@ipfw.ru>, =?UTF-8?Q?Olivier_Cochard=2DLabb=C3=A9?= <olivier@cochard.me>, "freebsd@intel.com" <freebsd@intel.com>, =?UTF-8?Q?Jev_Bj=C3=B6rsell?= <jev@sippysoft.com>, FreeBSD Net <freebsd-net@freebsd.org> Subject: Re: Poor high-PPS performance of the 10G ixgbe(9) NIC/driver in FreeBSD 10.1 Message-ID: <CAH7qZfvtT3VoqkZ8MROErkW9GsPg00Orc0GQ0R3kHpDTsacN%2BA@mail.gmail.com> In-Reply-To: <CAH7qZfvqBFtKfR97ZDhGhspKJ30=Mw=y-4UXOWRU_DXtJuyF6w@mail.gmail.com> References: <CAH7qZft-CZCKv_7E9PE%2B4ZN3EExhezMnAb3kvShQzYhRYb2jMg@mail.gmail.com> <77171439377164@web21h.yandex.ru> <55CB2F18.40902@FreeBSD.org> <CA%2BhQ2%2Bi37JzeUh8drxLSeeXHzYaRH9ZXvFyMBpF0XLHoiMSXMg@mail.gmail.com> <CAH7qZfvqBFtKfR97ZDhGhspKJ30=Mw=y-4UXOWRU_DXtJuyF6w@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
I think we are getting a better performance today with the IXGBE_FDIR switched off. It's not 100% decisive though, since we've only pushed it to little bit below 200kpps. We'll push more traffic tomorrow and see how it goes. -Maxim On Fri, Aug 14, 2015 at 10:29 AM, Maxim Sobolev <sobomax@freebsd.org> wrote= : > Hi guys, unfortunately no, neither reduction of the number of queues from > 8 to 6 nor pinning interrupt rate at 20000 per queue have not made any > difference. The card still goes kaboom at about 200Kpps no matter what. i= n > fact I've gone bit further, and after the first spike went on an pushed > interrupt rate even further down to 10000, but again no difference either= , > it still blows at the same mark. Although it did have effect on interrupt > rate reduction from 190K to some 130K according to the systat -vm, so tha= t > the moderation itself seems to be working fine. We will try disabling IXG= BE_FDIR > tomorrow and see if it helps. > > http://sobomax.sippysoft.com/ScreenShot391.png <- systat -vm with > max_interrupt_rate =3D 20000 right before overload > > http://sobomax.sippysoft.com/ScreenShot392.png <- systat -vm during issue > unfolding (max_interrupt_rate =3D 10000) > > http://sobomax.sippysoft.com/ScreenShot394.png <- cpu/net monitoring, > first two spikes are with max_interrupt_rate =3D 20000, the third one max= _interrupt_rate > =3D 10000 > > -Max > > On Wed, Aug 12, 2015 at 5:23 AM, Luigi Rizzo <rizzo@iet.unipi.it> wrote: > >> As I was telling to maxim, you should disable aim because it only matche= s >> the max interrupt rate to the average packet size, which is the last thi= ng >> you want. >> >> Setting the interrupt rate with sysctl (one per queue) gives you precise >> control on the max rate and (hence, extra latency). 20k interrupts/s giv= e >> you 50us of latency, and the 2k slots in the queue are still enough to >> absorb a burst of min-sized frames hitting a single queue (the os will >> start dropping long before that level, but that's another story). >> >> Cheers >> Luigi >> >> On Wednesday, August 12, 2015, Babak Farrokhi <farrokhi@freebsd.org> >> wrote: >> >>> I ran into the same problem with almost the same hardware (Intel X520) >>> on 10-STABLE. HT/SMT is disabled and cards are configured with 8 queues= , >>> with the same sysctl tunings as sobomax@ did. I am not using lagg, no >>> FLOWTABLE. >>> >>> I experimented with pmcstat (RESOURCE_STALLS) a while ago and here [1] >>> [2] you can see the results, including pmc output, callchain, flamegrap= h >>> and gprof output. >>> >>> I am experiencing huge number of interrupts with 200kpps load: >>> >>> # sysctl dev.ix | grep interrupt_rate >>> dev.ix.1.queue7.interrupt_rate: 125000 >>> dev.ix.1.queue6.interrupt_rate: 6329 >>> dev.ix.1.queue5.interrupt_rate: 500000 >>> dev.ix.1.queue4.interrupt_rate: 100000 >>> dev.ix.1.queue3.interrupt_rate: 50000 >>> dev.ix.1.queue2.interrupt_rate: 500000 >>> dev.ix.1.queue1.interrupt_rate: 500000 >>> dev.ix.1.queue0.interrupt_rate: 100000 >>> dev.ix.0.queue7.interrupt_rate: 500000 >>> dev.ix.0.queue6.interrupt_rate: 6097 >>> dev.ix.0.queue5.interrupt_rate: 10204 >>> dev.ix.0.queue4.interrupt_rate: 5208 >>> dev.ix.0.queue3.interrupt_rate: 5208 >>> dev.ix.0.queue2.interrupt_rate: 71428 >>> dev.ix.0.queue1.interrupt_rate: 5494 >>> dev.ix.0.queue0.interrupt_rate: 6250 >>> >>> [1] http://farrokhi.net/~farrokhi/pmc/6/ >>> [2] http://farrokhi.net/~farrokhi/pmc/7/ >>> >>> Regards, >>> Babak >>> >>> >>> Alexander V. Chernikov wrote: >>> > 12.08.2015, 02:28, "Maxim Sobolev" <sobomax@FreeBSD.org>: >>> >> Olivier, keep in mind that we are not "kernel forwarding" packets, >>> but "app >>> >> forwarding", i.e. the packet goes full way >>> >> net->kernel->recvfrom->app->sendto->kernel->net, which is why we hav= e >>> much >>> >> lower PPS limits and which is why I think we are actually benefiting >>> from >>> >> the extra queues. Single-thread sendto() in a loop is CPU-bound at >>> about >>> >> 220K PPS, and while running the test I am observing that outbound >>> traffic >>> >> from one thread is mapped into a specific queue (well, pair of queue= s >>> on >>> >> two separate adaptors, due to lagg load balancing action). And the >>> peak >>> >> performance of that test is at 7 threads, which I believe correspond= s >>> to >>> >> the number of queues. We have plenty of CPU cores in the box (24) wi= th >>> >> HTT/SMT disabled and one CPU is mapped to a specific queue. This >>> leaves us >>> >> with at least 8 CPUs fully capable of running our app. If you look a= t >>> the >>> >> CPU utilization, we are at about 10% when the issue hits. >>> > >>> > In any case, it would be great if you could provide some profiling >>> info since there could be >>> > plenty of problematic places starting from TX rings contention to som= e >>> locks inside udp or even >>> > (in)famous random entropy harvester.. >>> > e.g. something like pmcstat -TS instructions -w1 might be sufficient >>> to determine the reason >>> >> ix0: <Intel(R) PRO/10GbE PCI-Express Network Driver, Version - >>> 2.5.15> port >>> >> 0x6020-0x603f mem 0xc7c00000-0xc7dfffff,0xc7e04000-0xc7e07fff irq 40 >>> at >>> >> device 0.0 on pci3 >>> >> ix0: Using MSIX interrupts with 9 vectors >>> >> ix0: Bound queue 0 to cpu 0 >>> >> ix0: Bound queue 1 to cpu 1 >>> >> ix0: Bound queue 2 to cpu 2 >>> >> ix0: Bound queue 3 to cpu 3 >>> >> ix0: Bound queue 4 to cpu 4 >>> >> ix0: Bound queue 5 to cpu 5 >>> >> ix0: Bound queue 6 to cpu 6 >>> >> ix0: Bound queue 7 to cpu 7 >>> >> ix0: Ethernet address: 0c:c4:7a:5e:be:64 >>> >> ix0: PCI Express Bus: Speed 5.0GT/s Width x8 >>> >> 001.000008 [2705] netmap_attach success for ix0 tx 8/4096 rx >>> >> 8/4096 queues/slots >>> >> ix1: <Intel(R) PRO/10GbE PCI-Express Network Driver, Version - >>> 2.5.15> port >>> >> 0x6000-0x601f mem 0xc7a00000-0xc7bfffff,0xc7e00000-0xc7e03fff irq 44 >>> at >>> >> device 0.1 on pci3 >>> >> ix1: Using MSIX interrupts with 9 vectors >>> >> ix1: Bound queue 0 to cpu 8 >>> >> ix1: Bound queue 1 to cpu 9 >>> >> ix1: Bound queue 2 to cpu 10 >>> >> ix1: Bound queue 3 to cpu 11 >>> >> ix1: Bound queue 4 to cpu 12 >>> >> ix1: Bound queue 5 to cpu 13 >>> >> ix1: Bound queue 6 to cpu 14 >>> >> ix1: Bound queue 7 to cpu 15 >>> >> ix1: Ethernet address: 0c:c4:7a:5e:be:65 >>> >> ix1: PCI Express Bus: Speed 5.0GT/s Width x8 >>> >> 001.000009 [2705] netmap_attach success for ix1 tx 8/4096 rx >>> >> 8/4096 queues/slots >>> >> >>> >> On Tue, Aug 11, 2015 at 4:14 PM, Olivier Cochard-Labb=C3=A9 < >>> olivier@cochard.me> >>> >> wrote: >>> >> >>> >>> On Tue, Aug 11, 2015 at 11:18 PM, Maxim Sobolev < >>> sobomax@freebsd.org> >>> >>> wrote: >>> >>> >>> >>>> Hi folks, >>> >>>> >>> >>>> =E2=80=8BHi, >>> >>> =E2=80=8B >>> >>> >>> >>>> We've trying to migrate some of our high-PPS systems to a new >>> hardware >>> >>>> that >>> >>>> has four X540-AT2 10G NICs and observed that interrupt time goes >>> through >>> >>>> roof after we cross around 200K PPS in and 200K out (two ports in >>> LACP). >>> >>>> The previous hardware was stable up to about 350K PPS in and 350K >>> out. I >>> >>>> believe the old one was equipped with the I350 and had the >>> identical LACP >>> >>>> configuration. The new box also has better CPU with more cores >>> (i.e. 24 >>> >>>> cores vs. 16 cores before). CPU itself is 2 x E5-2690 v3. >>> >>> =E2=80=8B200K PPS, and even 350K PPS are very low value indeed. >>> >>> On a Intel Xeon L5630 (4 cores only) with one X540-AT2=E2=80=8B >>> >>> >>> >>> =E2=80=8B(then 2 10Gigabit ports)=E2=80=8B I've reached about 1.8M= pps >>> (fastforwarding >>> >>> enabled) [1]. >>> >>> But my setup didn't use lagg(4): Can you disable lagg configuratio= n >>> and >>> >>> re-measure your performance without lagg ? >>> >>> >>> >>> Do you let Intel NIC drivers using 8 queues for port too? >>> >>> In my use case (forwarding smallest UDP packet size), I obtain >>> better >>> >>> behaviour by limiting NIC queues to 4 (hw.ix.num_queues or >>> >>> hw.ixgbe.num_queues, don't remember) if my system had 8 cores. And >>> this >>> >>> with Gigabit Intel[2] or Chelsio NIC [3]. >>> >>> >>> >>> Don't forget to disable TSO and LRO too. >>> >>> >>> >>> =E2=80=8BRegards, >>> >>> >>> >>> Olivier >>> >>> >>> >>> [1] >>> >>> >>> http://bsdrp.net/documentation/examples/forwarding_performance_lab_of_a= n_ibm_system_x3550_m3_with_10-gigabit_intel_x540-at2#graphs >>> >>> [2] >>> >>> >>> http://bsdrp.net/documentation/examples/forwarding_performance_lab_of_a= _superserver_5018a-ftn4#graph1 >>> >>> [3] >>> >>> >>> http://bsdrp.net/documentation/examples/forwarding_performance_lab_of_a= _hp_proliant_dl360p_gen8_with_10-gigabit_with_10-gigabit_chelsio_t540-cr#re= ducing_nic_queues >>> >> _______________________________________________ >>> >> freebsd-net@freebsd.org mailing list >>> >> http://lists.freebsd.org/mailman/listinfo/freebsd-net >>> >> To unsubscribe, send any mail to "freebsd-net-unsubscribe@freebsd.or= g >>> " >>> > _______________________________________________ >>> > freebsd-net@freebsd.org mailing list >>> > http://lists.freebsd.org/mailman/listinfo/freebsd-net >>> > To unsubscribe, send any mail to "freebsd-net-unsubscribe@freebsd.org= " >>> _______________________________________________ >>> freebsd-net@freebsd.org mailing list >>> http://lists.freebsd.org/mailman/listinfo/freebsd-net >>> To unsubscribe, send any mail to "freebsd-net-unsubscribe@freebsd.org" >> >> >> >> -- >> -----------------------------------------+------------------------------= - >> Prof. Luigi RIZZO, rizzo@iet.unipi.it . Dip. di Ing. dell'Informazione >> http://www.iet.unipi.it/~luigi/ . Universita` di Pisa >> TEL +39-050-2217533 . via Diotisalvi 2 >> Mobile +39-338-6809875 . 56122 PISA (Italy) >> -----------------------------------------+------------------------------= - >> >> >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAH7qZfvtT3VoqkZ8MROErkW9GsPg00Orc0GQ0R3kHpDTsacN%2BA>