40Gbps http client benchmark

[Eugene Grosbein]

08.08.2016 8:57, Sepherosa Ziehau пишет:
> We have an optimized wrk here:
> https://github.com/sepherosa/wrk

I've glanced over your changes to wrk.
(Btw, you have minor bug there: missed 'N' short flag in getopt_long() invocation
for new --delay option, so "wrk --delay" works but "wrk -N" fails to parse command line.)

I managed to reproducible get 4.38GB/s (about 36Gbps) with your version of wrk
and it still eats over 60% of each core with "system time" using 8-core
Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz. Same speed I get without your changes.

I see that I cannot make even load for all four 10G ports of lagg interface
and that's obvious reason to be unable to saturate 40Gbps.

I need to create more connections for benchmark but then wrk eats even more CPU
and overall throughput decreases.