well, ovs-dpdk + userspace-tcp + app should be a better way. I just made this test so we can have a conclusion that ovs-dpdk + veth is not a good choice. And it worse than regular ovs.
I will do more performance testings on latest openvswitch and send a patch out ASAP. Thanks Zhenyu Gao 2017-06-01 10:16 GMT+08:00 Ben Pfaff <b...@ovn.org>: > Are you sure that this is the fastest way to interface OVS-DPDK to a > container? But, even if it is not, optimizations are welcome. > > On Thu, Jun 01, 2017 at 09:50:44AM +0800, Gao Zhenyu wrote: > > Here is the backgroud: > > > > I tried to consume ovs-dpdk(ovs-2.7.0, dpdk-16.11) for container-app and > > here is the topologic > > > > > > netserver > > |-------------------| > > | | > > | container | > > |--------veth----| > > | > > | |--------------------| > > |-------veth-| dpdk-ovs > > | netperf > > | > > | |-------------------| > > > > |---------dpdk----| | > > bare-metal | > > > > | > > --------------------- > > > > | | > > > > | | > > > > physical-nic--------------------------------------------physical-nic > > > > But the performance is worse than regular OVS and then I found sendmsg > cost > > 96% cpu cycles. Some packets were dropped due to insufficient cpu. > > So I tried to replace sendmsg with sendmmsg, it shows some performance > > improvement like: > > > > netperf -H 10.100.85.242 -t TCP_STREAM -l 60 > > 335.98Mb(sendmsg + ovs-dpdk) ----> 663Mb(sendmmsg + ovs-dpdk) > > > > (I turn off the veth's tx offloading because the dpdk would not do > > tx-checksum which introduces tcp-checksum error. ethtool -K eth1 tx > off) > > > > > > Thanks > > Zhenyu Gao > > > > > > 2017-05-31 23:41 GMT+08:00 Ben Pfaff <b...@ovn.org>: > > > > > On Wed, May 31, 2017 at 09:50:09AM +0800, Gao Zhenyu wrote: > > > > BTW, I would like to submit another patch to use sendmmsg to replace > > > > sendmsg. Sendmmsg get more benefit on throughput(my draft testing > show > > > 100% > > > > improvment). Do you think it is doable? > > > > > > I'm surprised that it makes a big difference, because I tested a > similar > > > change years ago and it did not. However, let's assume that it does. > > > In that case, of course we'd accept a change. It would be important to > > > retain support for older kernels and non-Linux kernels. > > > > _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
