[netperf-talk] Re: Difficultly in understanding the CPU utilization results for UDP stream test

[Rick Jones]

Brijraj Vaghani wrote:
> Rick,
> 
> Did you try this out on a Virtual Linux machine or a dedicated Linux 
> machine?

This was on a dedicated Linux machine.  I must have missed the bit where 
you said that you were running on a virtual machine.  If you are running 
on a linux virtual machine, I'm not at all surprised that netperf might 
get the CPU util wrong, it knows nothing about virtual machines.  Then 
again, I'm not sure if top knows all it needs to know either.  Either 
way, I suspect that netperf will need to be taught quite a bit to get 
good CPU util in a virtual machine environment.

And that of course raises a meta issue - exactly how and which CPU 
utilization should netperf (try to) report when running on a virtual 
client?  Should it be the overall CPU util of the hypervisor?  To just 
report the "CPU util" of the guest OS would be rather like just using 
the per-process CPU util of netperf itself - there will be (potentially 
lots of) CPU consumed elsewhere which will not be charged to the 
process/guest.


WRT getting onto later versions of netperf, the "best" way would be to 
use a subversion client and either grab a tagged version from:

http://www.netperf.org/svn/netperf2/tags/

or the top of trunk from:

http://www.netperf.org/svn/netperf2/trunk/

There are some tar and/or zip files on:

ftp://ftp.netperf.org/

as well.

rick jones

> 
> Thanks,
> Brijraj
> 
> On 7/3/07, Rick Jones <rick.jones2 at hp.com> wrote:
> 
>> Well, I've tried to reproduce your results, with a pair of four-core
>> systems running a 2.6.9.mumble RHAS4u5 environment and the top of trunk.
>>     With --enable-intervals, I see "sane" CPU util numbers which appear
>> to match the output of top.  I am using command-lines like:
>>
>> [root at hpcpc107 netperf2_work]# src/netperf -t UDP_STREAM -H hpcpc106 -C
>> -b 8 -w 10000 -l 120
>> UDP UNIDIRECTIONAL SEND TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
>> hpcpc106.cup.hp.com (16.89.84.106) port 0 AF_INET : interval
>> Socket  Message  Elapsed      Messages                   CPU      Service
>> Size    Size     Time         Okay Errors   Throughput   Util     Demand
>> bytes   bytes    secs            #      #   10^6bits/sec % US     us/KB
>>
>> 135168   65507   119.94         96      0        0.4     -1.00    -1.000
>> 135168           119.94         96               0.4     0.11     88.518
>>
>> and I am _not_ using a -n option.  So, unless we have some other
>> command-line differences, or there is something in an older netperf
>> version (this is top of trunk, which will be 2.4.4 when I decide to tag
>> it) it seems there is something truly odd in your systems.
>>
>>
>> rick
>>