[netperf-talk] concurrent sessions with Netperf 2.4.2

[Rick Jones]

anil mishra wrote:
> Rick,
> I did following:
> for i in 1 2 3 4
> do
> netperf -t TCP_STREAM -H 192.168.x.x -p 99 -c -C
> netperf -t TCP_RR -H 192.168.x.x
> for j in 1
> do
> netperf -t TCP_STREAM -H 192.168.x.x -c -C
> netperf -t TCP_MAERTS -H 192.168.x.x -c -C
> done
> done

Since there are no '&' at the ends of the netperf command lines, the 
shell will run them one at a time rather than run them in parallel in 
the background.

Further, one should _not_ trust this concurrent session method with only 
one iteration of each test and such a short run time.  The potential for 
startup/shutdown skew error is too great.

Per the online docs :) one should use the confidence intervals mechanism 
to get netperf to run more than one iteration for each invokation.  So, 
adding a -i 10 or -i 30 to each netperf command line would be indicated. 
  Might not be a bad idea to increase the individual run times also with 
say -l 30.  A combination of -l 30 and -i 30 means the run will go for 
30*30 = 900 seconds ie 15 minutes.

Any particular reason you are sending the first TCP_STREAM test to a 
different control port number with -p?

Unless you like parsing jumbled-up output, you might also want to use -P 
0 to turn-off the test banners.  If you want identification of 
individual results, the 2.4.3 (not sure if 2.4.2 has it) version has the 
-B option to add the passed-in text to the brief output, on the same 
line so it sticks with the rest of it.

Of course, the better place to be running concurrent tests is netperf4, 
even if the interface is still a bit unwieldy.  Netperf4 has support for 
avoiding skew errors entirely.

happy benchmarking,

rick jones