[netperf-talk] Urgent UDP_Stream question

[Rick Jones]

On 11/10/2011 08:06 AM, Andrew Gallatin wrote:
>
>> UDP_STREAM test:
>> Socket Size, bytes Message Size, bytes Elapsed Time, secs Messages
>> Throughput, 10^6bits/sec
>> Okay Errors
>> 112640 65507 60.00 57408 0 501.42
>> 112640 60.00 94 0.82
>
> A big part of the problem here is that netperf has been asked
> (implicitly) to send 64KB datagrams. Assuming a standard 1500b
> MTU, these will get fragmented into around 40 separate IP
> fragments per datagram. If even just a single fragment is
> dropped, then the entire message is lost.

It would be worthwhile to check the statistics on the receiver - both 
for UDP to see if there were socket buffer overflows, and for IP to 
check for those fragment drops, which should manifest as IP fragment 
reassembly failures (though thanks to the 16 bit IP fragment ID they can 
manifest as UDP checksum failures - web search for "frankengram."