docs: even more benchmarks

[mreiferson]

#438 got the ball rolling, but I'd like to add a few more benchmark modes:

• varying numbers of topics/channels
• varying numbers of producers/consumers
• varying size of message
• latency measurements
• writing to diskqueue

[thomasmodeneis]

Hi @mreiferson
I really appreciate the work you are doing on the performance benchmarking thing and doing fair comparisons, not like that blog post (Mac 2.6GB...).

I'm using NSQ for couple months and doing migration from the actual message system I use (Redis/Lua) to NSQ and I find this issue really important and I would like to help.

On the issue #438 you mentioned the results of your benchmarking, can you please detail what each one of the columns for the results are for? Oh and the script you used to bench was this one? https://github.com/tylertreat/mq-benchmarking

[I 140916 15:41:29 bench:239] [bench_reader] 10.514560s - 206.496000mb/s - 1082649.021000ops/s - 9.711882us/op

Thank you

[mreiferson]

@thomasmodeneis this sounds great, thanks for the help!

The script I used is the one in this repo (that was merged in #438) bench/bench.py.

The columns are (in order) total time, throughput in megabytes per second, number of total operations (messages), and then total time / ops.

The last value (total time / ops) is misleading as a result of "coordinated omission" (see http://www.azulsystems.com/sites/default/files/images/HowNotToMeasureLatency_LLSummit_NYC_12Nov2013.pdf) and should be improved.

[mreiferson]

moving this issue to https://github.com/nsqio/nsqio.github.io