Socketplane performance and steadibility?

[resouer]

I did a simple benchmark test by using qperf, to my surprise, sockeplane did no play well.

SocketPlane container to container:
tcp_bw:
bw = 6.7 MB/sec
tcp_lat:
latency = 290 us

Host to host:
tcp_bw:
bw = 11.7 MB/sec
tcp_lat:
latency = 141 us

It's not as good as flannel+vxlan which is 70%-80% of "native", according to my test.

Does anyone has historical benchmark test results to share & comparison with flannel?

[nerdalert]

• Last one I remember us doing was November of last year:
  Proposal: Native Docker Multi-Host Networking moby/moby#8951 (comment)
• The environment @dave-tucker used is here: https://github.com/dave-tucker/overlay-comparison
• The socketplane code doesn't inject itself in the datapath in any way other then setting up the environment. OVS versions are relevant particularly prior to flow wildcarding (v1.9ish).
• Here are some DPDK numbers Martin posted recently if interested in speeds and feeds. https://twitter.com/martin_casado/status/628567114169884672

[resouer]

Newly tested again, I tore down the former env in vsphere and build a freshly new test cluster by vagrant, here's the result:

container to container:
tcp_bw:
bw = 168 MB/sec
tcp_lat:
latency = 102 us

host to host:
tcp_bw:
bw = 245 MB/sec
tcp_lat:
latency = 88.4 us

This time, not bad. But not the as good as I expected (see https://groups.google.com/forum/#!topic/rancherio/cIFacrkQ7jA/ ). I think we lost performance during vxlan tunnel?

Another issue is that during my test, socketplane did not show steadily, for example, containers can ping to each other one time and failed another time. It seems socketplane requires a very 'clean' env to run and easy to fail otherwise.