[netperf-talk] Help me to understand the results

[Rick Jones]

Steevan Rodrigues wrote:
> Hi Rick,
> 
> Thanks for clarification.
> How to change the method used to calculate CPU utilization ?
> Do I need to configure netperf with flag enable-cpuutil=method ?
> I am actually using this netperf on ARM board.
> Which is the best method ?

The configure script will attempt to pick the correct method on its own.  For 
"linux" /proc/stat is the "correct" method, so there shouldn't be a need to use 
the --enable-cpuutil=method bit.

IIRC, ARM is not HW threaded and is a small number of cores.  I would have 
expected /proc/stat to work there, but then Linux means never having to say you 
are finished with debugging :)

At this point, the output of netperf -d and of /proc/stat would be the way to go 
- making certain that netperf and the /proc/stat format are "in sync"

A snippet of top output showing the discrepancy would be goodness too.

happy benchmarking,

rick jones

> 
> Regards,
> Steev
> 
> 
> On Thu, Jan 27, 2011 at 11:36 PM, Rick Jones <rick.jones2 at hp.com 
> <mailto:rick.jones2 at hp.com>> wrote:
> 
>     [discussions started in netperf-talk should remain there... ]
> 
>     Steevan Rodrigues wrote:
> 
>         Hi Rick,
> 
>         Thanks for the explanation. That really helped us.
> 
>         I have one more question on CPU utilization reported by netperf.
> 
>         Does netperf report CPU utilization for the "netperf" tool or is
>         it entire CPU utilization accounting for other tasks that are
>         running ?
> 
>         I am asking this question because I am seeing large difference
>         between CPU utilization reported by netperf and the "top" command.
> 
>         Could you please clarify this?
> 
> 
>     Sadly,
>     http://www.netperf.org/svn/netperf2/trunk/doc/netperf.html#CPU-Utilization
>     does not address that directly.  I'll have to update it at some point.
> 
>     Netperf believes it is measuring the CPU utilization of the entire
>     system and ass-u-me-s it is the only source of that CPU utilization.
>       That said, with the advent of HW threading and such, measuring
>     "CPU utilization" has become rather more complicated and depending
>     on the system, and the method netperf is using (see the URL above) I
>     could see where there might be discrepancies.
> 
>     Based on the previous emails, and the netperf output contained
>     therein, I would guess that the method being used by netperf is the
>     /proc/stat method.  (The 'S' in the header portion for CPU
>     utilization is the thing that makes me guess that).  There have been
>     some "issues" in the past with format changes in /proc/stat that
>     were not noticed by netperf.  You might compare the output of
>     /proc/stat with the source code in src/netcpu_procstat.c  While
>     doing-so, adding a -d to the global command line options to get the
>     code to be a bit chatty about what it is doing might be goodness.
> 
>     happy benchmarking,
> 
>     rick jones
> 
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> netperf-talk mailing list
> netperf-talk at netperf.org
> http://www.netperf.org/cgi-bin/mailman/listinfo/netperf-talk