[netperf-talk] CPU Utilization Issues

[Rick Jones]

Lentz, Benjamin A. wrote:
> Our vendor is reporting that in request-response tests, our IBM 3850
> system is running about half as fast as their small Core 2 Duo system.
> Any idea what would cause these values to be reported? I'm not sure what
> to make of this.
> 
> Thanks again in advance!

I'm going to have to insist you at least by a ProCurve switch or three 
after all this :) :) :)

Anyhow, latency, especially small packet latency, is dominated by 
path-length vs CPU "oomph" and the interrupt coalescing decisions made 
by the NIC/driver.  An older writeup at:

ftp://ftp.cup.hp.com/dist/networking/briefs/nic_latency_vs_tput.txt

discusses the latter.

There can also be an effect from where the interrupt is serviced 
relative to where the IO is connected to the fabric.

Finally, while I suspect it is not happening here, netperf makes no 
attempt to recover from lost UDP datagrams - this means if a UDP 
datagram for either a request or a reponse is lost, the test will come 
to a screeching halt until the test timer expires.  So, when one sees an 
especially low UDP_RR result, rerunning as a TCP_RR test is a good idea.

One of these days I really need to make measuring CPU utilization the 
default on those systems which do not require calibration...  in the 
meantime, please get into the habit of using -c and -C, and strongly 
encourage your vendor to do the same.


> On our vendor's smaller system:
> 
> [root at es3rac-1 src]# sh udp_rr_script 192.168.2.102
> 
> ------------------------------------------------------
> Testing with the following command line:
> ./netperf -l 60 -H 192.168.2.102 -i 10,2 -I 99,10 -t UDP_RR -- -r 1,1 -s
> 0 -S 0
> 262144 4194304 1        1       60.00    12763.70
> 262144 4194304
> 
> ------------------------------------------------------
> Testing with the following command line:
> ./netperf -l 60 -H 192.168.2.102 -i 10,2 -I 99,10 -t UDP_RR -- -r 64,64
> -s 0 -S 0
> 262144 4194304 64       64      60.00    12190.79
> 262144 4194304
> 
> ------------------------------------------------------
> Testing with the following command line:
> ./netperf -l 60 -H 192.168.2.102 -i 10,2 -I 99,10 -t UDP_RR -- -r
> 100,200 -s 0 -S 0
> 262144 4194304 100      200     60.00    11546.24
> 262144 4194304
> 
> ------------------------------------------------------
> Testing with the following command line:
> ./netperf -l 60 -H 192.168.2.102 -i 10,2 -I 99,10 -t UDP_RR -- -r
> 1024,1024 -s 0 -S 0
> 262144 4194304 1024     1024    60.00    7662.73
> 262144 4194304
> 
> If you wish to submit these results to the netperf database at
> http://www.cup.hp.com/netperf/NetperfPage.html, please submit each
> datapoint individually. Individual datapoints are separated by
> lines of dashes.
> 
> 
> And from our larger system:
> 
> [root at ykrhappsch01 src]# sh udp_rr_script 192.168.101.11
> 
> ------------------------------------------------------
> Testing with the following command line:
> ./netperf -l 60 -H 192.168.101.11 -i 10,2 -I 99,10 -t UDP_RR -- -r 1,1
> -s 0 -S 0
> 262144 4194304 1        1       60.00    7780.14
> 262144 4194304
> 
> ------------------------------------------------------
> Testing with the following command line:
> ./netperf -l 60 -H 192.168.101.11 -i 10,2 -I 99,10 -t UDP_RR -- -r 64,64
> -s 0 -S 0
> 262144 4194304 64       64      60.00    7434.79
> 262144 4194304
> 
> ------------------------------------------------------
> Testing with the following command line:
> ./netperf -l 60 -H 192.168.101.11 -i 10,2 -I 99,10 -t UDP_RR -- -r
> 100,200 -s 0 -S 0
> 262144 4194304 100      200     60.00    7035.14
> 262144 4194304
> 
> ------------------------------------------------------
> Testing with the following command line:
> ./netperf -l 60 -H 192.168.101.11 -i 10,2 -I 99,10 -t UDP_RR -- -r
> 1024,1024 -s 0 -S 0
> 262144 4194304 1024     1024    60.00    4494.62
> 262144 4194304
> 
> If you wish to submit these results to the netperf database at
> http://www.cup.hp.com/netperf/NetperfPage.html, please submit each
> datapoint individually. Individual datapoints are separated by
> lines of dashes.

Hmm, I guess I better update some of those ancient scripts :)  Or, get a 
new netperf database online...

happy benchmarking,

rick jones

> _______________________________________________
> netperf-talk mailing list
> netperf-talk at netperf.org
> http://www.netperf.org/cgi-bin/mailman/listinfo/netperf-talk