[zeromq-dev] perf/throughput tests calculates bandwith wrong

[Erik Hugne]

Hi.
I'm currently creating a transport module for the TIPC protocol for 0MQ.
When doing benchmarks vs TCP, i've noticed some odd behavior for higher buffer sizes (above 4096B).
Graph comparing TIPC/TCP throughput (bundled perf test).
http://imgur.com/fwA4gDV

Rather than posting all benchmark numbers here, i've uploaded them as gists.
perf/thr run for packet sizes 4096-65536, buffer increments of 512b.
https://gist.github.com/Hugne/5028836

This does not occur every time, for example, here's two consecutive tests with buffer size 8192:
hp1-blade9:~/zeromq3-x/perf # ./local_thr tcp://eth1:5555 8192 1000000
message size: 8192 [B]
message count: 1000000
mean throughput: 53544 [msg/s]
mean throughput: 3509.060 [Mb/s]

Hmm, 8192*53544 = 438632448 bytes per second ~=  413 Mb/s..
3509Mb/s.. 27 Gigabit/s.. i dont think so..

hp1-blade9:~/zeromq3-x/perf # ./local_thr tcp://eth1:5555 8192 1000000
message size: 8192 [B]
message count: 1000000
mean throughput: 74661 [msg/s]
mean throughput: 598.016 [Mb/s]

598.. This is more normal figures, which matches the iperf3 measurements @ 8K (523.52 Mb/s, 4.09Gbits).

Iperf3 run with buffer increments of 1k (smallest unit is 1k):
https://gist.github.com/Hugne/5028859


//E