Hi,

<mailto:[email protected]>
    Trying my luck see if you have seen similar issue

While enabled multiqueue settings in qemu to bring up 2 virtual machines and 
doing netperf test.
The traffic was not evenly distributed for IPv4 FRAGMENT packets(i confirmed 
that the issue is not observed on non-fragment traffic)

qemu version 5.0.0
openvswitch 2.13.5 running kernal datapath
openvswitch.ko 4.18

-netdev type=tap,id=net1,vhost=on,ifname=vm1,queues=6 -device 
virtio-net-pci,host_tso4=on,host_tso6=on,netdev=net1,id=net1,mq=on,vectors=13

-netdev type=tap,id=net2,vhost=on,ifname=vm2,queues=6 -device 
virtio-net-pci,host_tso4=on,host_tso6=on,netdev=net2,id=net2,mq=on,vectors=13

and i added the 2 interfaces to ovs bridge

on host ethtool -i vm1
  driver: tun
  version: 1.6
on host VM(ubuntu20.04)  ethtool -i eth1
  driver:virtio-net
  version:1.0.0

i could see the 6 queues in /sys/class/net/vm1/queues on host as well as guest.

The tricky thing is, while doing 5 parallel netperf process from vm1 to 
vm2(hitting kernel datapath as well as heavily use ovs-vswitchd for flow table 
lookup), in the final result, i could see 90%+ connection counters was on one 
single netperf process

"netperf -H 1.1.1.2 -p 1810 -t UDP_RR -R 1 -r 8K,8K"
"netperf -H 1.1.1.2 -p 1811 -t UDP_RR -R 1 -r 8K,8K"
"netperf -H 1.1.1.2 -p 1812 -t UDP_RR -R 1 -r 8K,8K"
"netperf -H 1.1.1.2 -p 1813 -t UDP_RR -R 1 -r 8K,8K"
"netperf -H 1.1.1.2 -p 1814 -t UDP_RR -R 1 -r 8K,8K"

Wonder if you experienced anything related to this.

My search seems reveal that multiqueue settings apply mostly to ovs dpdk with 
pmd port queuing settings.
not sure if there is any known issue related with kernel mode OVS.


test i did:

  1.  cat /proc/interrupts in guest vm, it seems the interrupt counters are 
evenly distributed

  1.  disable multiqueue, use one single queue, remove 
"queues,..,mq=on,vectors=13" open, then fragments get evenly distributed,
  2.  however, in this case, tcp_rr performance drops by 30%
  3.
  4.  add vhostforce option, no benefit

many thanks
eddy
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to