[netperf-talk] Negative number of message Okay in UDP_STREAM Test
Rick Jones
rick.jones2 at hp.com
Fri Dec 7 10:18:08 PST 2012
On 12/06/2012 10:40 PM, Piyush Verma wrote:
> Here I found the output of UDP_STREAM test for 24 hrs and getting
> negative number of message Okay packets-
>
> [root at rhel61-srv-64 vermap]# ./netperf -H 172.16.128.155 -p 49152 -t
> udp_stream -l 86400 -- -m 8192 -s 32768 -S 65536 &
> [1] 4115
> UDP UNIDIRECTIONAL SEND TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
> 172.16.128.155 (172.16.128.155) port 0 AF_INET
>
> Socket Message Elapsed Messages
> Size Size Time Okay Errors Throughput
> bytes bytes secs # # 10^6bits/sec
>
> 65536 8192 86400.00 -2081184671 0 1679.20
> 131072 86400.00 -2081420383 1679.02
>
> Please let me know what does negative message Okay means here?
It means that the likely 32 bit signed value being used to count message
has "wrapped." That is, in that 24 hour period netperf sent over ~2
billion messages.
Given that I don't see "MIGRATED" in the test banner, I'm guessing the
(unstated) netperf version is earlier than 2.5.0. If having a negative
value there is a problem, you might give top-of-trunk a try, where the
counters for the number of send calls have been bumped to 64 bit.
However, looking at the definition for what is returned by the netserver
side, that one can still wrap around - it is an unsigned 32-bit quantity.
Netperf wasn't really ever expected to have a single instance run that
long. Typically 24 hours for a benchmark is considered a tad excessive.
Of course, some folks (ab)use netperf for functional QA purposes - 24
CHO and such. In that case, if negative values are an issue, and 24
hours of operation under one netperf command line is still required, I
might suggest (ab)using the confidence intervals. Set the width quite
wide and the confidence low, and then make the number of iterations the
maximum - 30 - with the run time 86400/30 or 2880:
netperf -H 172.16.128.155 -p 49152 -t udp_stream -I 95,50 -i 30,30 -l
2880 -m 8192 -s 32768 -S 65536
Or, just run a series of shorter-length netperfs in a row.
happy benchmarking,
rick jones
More information about the netperf-talk
mailing list