[netperf-talk] global question concerning Netperf test and SMP support

Thu Apr 26 10:47:34 PDT 2012

On 04/26/2012 07:19 AM, Simon Duboue wrote:
> Hello everyone,
>
> Here are a part of my last results:
>
> netperf -l 60 -H 10.0.17.200 -t TCP_STREAM -cC -i 10,2 -I 99,5 -- -m
> 32768 -s 57344 -S 57344
> netperf -l 60 -H 10.0.17.200 -t TCP_STREAM -cC -i 10,2 -I 99,5 -- -m
> 32768 -s 57344 -S 57344
> netperf -l 60 -H 10.0.17.200 -t TCP_STREAM -cC -i 10,2 -I 99,5 -- -m
> 32768 -s 57344 -S 57344
> netperf -l 60 -H 10.0.17.200 -t TCP_STREAM -cC -i 10,2 -I 99,5 -- -m
> 32768 -s 57344 -S 57344
> netperf -l 60 -H 10.0.17.200 -t TCP_STREAM -cC -i 10,2 -I 99,5 -- -m
> 32768 -s 57344 -S 57344

These were consecutive (one at a time) right?  Because if they were 
concurrent (all at once) you really must set max,min iterations for the 
confidence intervals to the same value, lest one or more of the 
instances terminate before the others.

> TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.17.200
> (10.0.17.200) port 0 AF_INET : +/-2.5% @ 99% conf. : demo
> Recv Send Send Utilization Service Demand
> Socket Socket Message Elapsed Send Recv Send Recv
> Size Size Size Time Throughput local remote local remote
> bytes bytes bytes secs. 10^6bits/s % S % S us/KB us/KB
>
> 114688 114688 32768 60.00 1471.34 25.67 33.66 2.889 15.081
> TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.17.200
> (10.0.17.200) port 0 AF_INET : +/-2.5% @ 99% conf. : demo
> Recv Send Send Utilization Service Demand
> Socket Socket Message Elapsed Send Recv Send Recv
> Size Size Size Time Throughput local remote local remote
> bytes bytes bytes secs. 10^6bits/s % S % S us/KB us/KB
>
> 114688 114688 32768 60.00 1456.63 25.68 33.67 2.901 15.043
> TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.17.200
> (10.0.17.200) port 0 AF_INET : +/-2.5% @ 99% conf. : demo
> Recv Send Send Utilization Service Demand
> Socket Socket Message Elapsed Send Recv Send Recv
> Size Size Size Time Throughput local remote local remote
> bytes bytes bytes secs. 10^6bits/s % S % S us/KB us/KB
>
> 114688 114688 32768 60.01 1512.94 25.68 33.67 2.816 14.969
> TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.17.200
> (10.0.17.200) port 0 AF_INET : +/-2.5% @ 99% conf. : demo
> Recv Send Send Utilization Service Demand
> Socket Socket Message Elapsed Send Recv Send Recv
> Size Size Size Time Throughput local remote local remote
> bytes bytes bytes secs. 10^6bits/s % S % S us/KB us/KB
>
> 114688 114688 32768 60.03 1465.22 25.67 33.68 2.920 14.895
> TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.17.200
> (10.0.17.200) port 0 AF_INET : +/-2.5% @ 99% conf. : demo
> Recv Send Send Utilization Service Demand
> Socket Socket Message Elapsed Send Recv Send Recv
> Size Size Size Time Throughput local remote local remote
> bytes bytes bytes secs. 10^6bits/s % S % S us/KB us/KB
>
> 114688 114688 32768 60.03 1498.52 25.67 33.68 2.819 14.849
>
> Total throughput = 1471.34 + 1456.63 + 1512.94 + 1465.22 + 1498.52 =
> 7404.65 Mbits/s.

Did you "fix-up" the output?  If they were indeed concurrent (all 
launched in the background) I would have expected to see all the test 
banners, then all the test results.  By way of quick-and-dirty example:

raj at tardy:~/netperf2_trunk$ for i in `seq 1 5`
 > do
 > netperf -H 192.168.1.3 -l 10 &
 > done
[1] 11289
[2] 11290
[3] 11291
[4] 11292
[5] 11293
raj at tardy:~/netperf2_trunk$ MIGRATED TCP STREAM TEST from 0.0.0.0 
(0.0.0.0) port 0 AF_INET to 192.168.1.3 () port 0 AF_INET : demo
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 
192.168.1.3 () port 0 AF_INET : demo
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 
192.168.1.3 () port 0 AF_INET : demo
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 
192.168.1.3 () port 0 AF_INET : demo
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 
192.168.1.3 () port 0 AF_INET : demo
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

  87380  16384  16384    10.39       1.83
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

  87380  16384  16384    10.40       1.85
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

  87380  16384  16384    10.42       1.85
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

  87380  16384  16384    10.43       1.99
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

  87380  16384  16384    10.46       1.93

The output as you have presented it makes it look as though the tests 
ran one-at-a-time.

Another sanity check would be to compare the service demands between a 
single (all by itself) test and then with your five-at-once tests. 
Since each netperf/netserver measures overall CPU utilization, but only 
what it itself transferred, when you run actually concurrent tests, the 
service demands reported by the ostensibly-all-at-once tests will be 
(much) higher (and wrong :).  If you see the same service demand for 
each of the ostensibly-all-at-once as you do for a single test, you know 
the ostensibly-all-at-once were not really all-at once.

> The CPU utilization seems correct and are very close together and they
> are not negative! (this seems to appear more with UDP stream test than
> with TCP ones)
> I observe that these negative values mainly appears on the first results
> of my test script.
>
> If you have any remarks.

Apart from what I've already mentioned, I'll simply repeat my concern 
about the negative CPU utilization reported before and leave it at that.

rick