> -----Original Message----- > From: Ilya Maximets [mailto:i.maxim...@samsung.com] > Sent: Tuesday, February 20, 2018 2:55 PM > To: ovs-dev@openvswitch.org; O Mahony, Billy <billy.o.mah...@intel.com> > Subject: Re: Re: [ovs-dev] [RFC 0/2] Ingress Scheduling > > > This patch set implements the 'preferential read' part of the feature > > of ingress scheduling described at OvS 2017 Fall Conference > > https://www.slideshare.net/LF_OpenvSwitch/lfovs17ingress-scheduling- > 82280320. > > > > It allows configuration to specify an ingress priority for and entire > > interface. This protects traffic on higher priority interfaces from > > loss and latency as PMDs get overloaded. > > > > Results so far a are very promising; For a uniform traffic > > distribution as total offered load increases loss starts on the lowest > > priority port first and the highest priority port last. > > > > When using four physical ports with each port forwarded to one of the > > other ports. The following packets loss is seen. The EMC was bypassed > > in this case and a small delay loop was added to each packet to > > simulate more realistic per packet processing cost of 1000cycles approx. > > > > Port dpdk_0 dpdk_1 dpdk_2 dpdk_3 > > Traffic > > Dist. 25% 25% 25% 25% > > Priority 0 1 2 3 > > n_rxq 8 8 8 8 > > > > Total > > Load Kpps Loss Rate Per Port (Kpps) > > 2110 0 0 0 0 > > 2120 5 0 0 0 > > 2920 676 0 0 0 > > 2930 677 5 0 0 > > 3510 854 411 0 0 > > 3520 860 415 3 0 > > 4390 1083 716 348 0 > > 4400 1088 720 354 1 > > > > > > Even in the case where most traffic is on the priority port this > > remains the > > case: > > > > Port dpdk_0 dpdk_1 dpdk_2 dpdk_3 > > Traffic > > Dist. 10% 20% 30% 40% > > Priority 0 1 2 3 > > n_rxq 8 8 8 8 > > > > Total > > Load Kpps Loss Rate Per Port (Kpps) > > 2400 0 0 0 0 > > 2410 5 0 0 0 > > 2720 225 5 0 0 > > 2880 252 121 9 0 > > 3030 269 176 82 3 > > 4000 369 377 384 392 > > 5000 471 580 691 801 > > > > The latency characteristics of the traffic on the higher priority > > ports is also protected. > > > > Port dpdk_0 dpdk_1 dpdk_2 dpdk_3 > > Traffic > > Dist. 10% 20% 30% 40% > > Priority 0 1 2 3 > > n_rxq 8 8 8 8 > > > > Total dpdk0 dpdk1 dpdk2 dpdk3 > > Load Kpps > > 2400 113 122 120 125 > > 2410 36117 571 577 560 > > 2720 323242 14424 3265 3235 > > 2880 391404 33508 10075 4600 > > 3030 412597 35450 17061 7965 > > 4000 414729 36070 17740 11106 > > 5000 416801 36445 18233 11653 > > > > Some General setup notes: > > Fortville. (X710 DA4. firmware-version: 6.01 0x800034af 1.1747.0) > > Intel(R) Xeon(R) CPU E5-2695 v3 @ 2.30GHz One pmd Port fwding port > > 0<->1, 2 <-> 3 Frame 64B, UDP 221 streams per port. > > OvS base - 4c80644 http://github.com/istokes/ovs dpdk_merge_2_9. Added > 600cycles approx pkt processing in order to bring per packet cost to ~1000 > cycles. > > DPDK v17.11.1 > > > > Hi Billy, thanks for your work on this feature. > I have one question here. Have you tasted heterogeneous configurations?
[[BO'M]] I have performed some tests for virtual devices which shows a similar effect. VMs with prioritized vhost ports achieve higher tcp transfer at the cost of b/w between non-prioritized. But I have not tested a mixture of either physical & virtual or different types of NIC. I do plan such tests as part of a closer characterization of how TCP transfers are affected. > I mean situations where PMD thread polls different types of ports (NICs). > It's known that RX operations has different cost for different port types > (like > virtual and physical) and also for different hardware NICs because of > different > implementations of DPDK PMD drivers. [[BO'M]] That is an interesting point. Any scheme that protects (ie reads more frequently) from an expensive-to-read port will have to have a greater cost on the remaining ports than if it were protecting a cheap-to-read port . Some packets are always going to be more costly than other depending on how they are processed but also as you point out on their ingress and egress port types. And if you want to process one more 'expensive' packets it's going to cost you more than one cheaper packet. The different levels of prioritization do offer a little flexibility here effectively allowing the spending of some extra cycles or a lot of extra cycles to protect traffic on different ports. > > So, my concern: Is it possible that setting higher priority for a slower port > type > (maybe vhost) may degrade overall performance of current PMD thread. > It'll be nice to have this kind of testing results. > > Best regards, Ilya Maximets. > > > Billy O'Mahony (2): > > ingress scheduling: schema and docs > > ingress scheduling: Provide per interface ingress priority > > > > Documentation/howto/dpdk.rst | 18 ++++++++++++ > > include/openvswitch/ofp-parse.h | 3 ++ > > lib/dpif-netdev.c | 47 +++++++++++++++++++++--------- > > lib/netdev-bsd.c | 1 + > > lib/netdev-dpdk.c | 64 > +++++++++++++++++++++++++++++++++++++++-- > > lib/netdev-dummy.c | 1 + > > lib/netdev-linux.c | 1 + > > lib/netdev-provider.h | 11 ++++++- > > lib/netdev-vport.c | 1 + > > lib/netdev.c | 23 +++++++++++++++ > > lib/netdev.h | 2 ++ > > vswitchd/bridge.c | 2 ++ > > vswitchd/vswitch.ovsschema | 9 ++++-- > > vswitchd/vswitch.xml | 40 ++++++++++++++++++++++++++ > > 14 files changed, 205 insertions(+), 18 deletions(-) > > > > -- > > 2.7.4 _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev