[Intel-wired-lan] [E1000-devel] Fragmented UDP packets trigger rx_missed_errors on 82599EB

[Tantilov, Emil S]

>-----Original Message-----
>From: Fan Du [mailto:fengyuleidian0615 at gmail.com] 
>Sent: Monday, April 06, 2015 6:52 PM
>Subject: Re: [E1000-devel] Fragmented UDP packets trigger rx_missed_errors on 82599EB
>
>于 2015年04月03日 01:52, Tantilov, Emil S 写道:
>>> >
>>> ># numactl --hardware
>>> >available: 4 nodes (0-3)
>>> >node 0 cpus: 0 1 2 3 4 20 21 22 23 24
>>> >node 0 size: 24466 MB
>>> >node 0 free: 22444 MB
>>> >node 1 cpus: 5 6 7 8 9 25 26 27 28 29
>>> >node 1 size: 16384 MB
>>> >node 1 free: 15831 MB
>>> >node 2 cpus: 10 11 12 13 14 30 31 32 33 34
>>> >node 2 size: 16384 MB
>>> >node 2 free: 15791 MB
>>> >node 3 cpus: 15 16 17 18 19 35 36 37 38 39
>>> >node 3 size: 24576 MB
>>> >node 3 free: 22508 MB
>>> >node distances:
>>> >node 0 1 2 3
>>> >0: 10 21 31 31
>>> >1: 21 10 31 31
>>> >2: 31 31 10 21
>>> >3: 31 31 21 10
>>> Since you have 4 nodes you may want to check your board layout and try to pin the queues and iperf to the same
>>   node as the network interface. See if that helps.
>
>Thanks for the hints.
>
>UDP is used here, so it actually doesn't matter to tie iperf server on the same core for receiving the flow.
>I didn't launch iperf server, only running iperf client to send *fragmented* UDP packets.
>On the server side, still ifconfig shows dropped packets increasing, and ethtool -S confirmed with
>rx_missed_errors also climbing.

One difference is that fragmented UDP packets would go to a single queue, provided you have UDP RSS enabled. 

You can force missed packets by overwhelming the receiver with small UDP packets, so I'm not sure that the UDP packets being fragmented have much to do with it.

>Client: iperf -c SEVER_IP -u  -b 10G -i 1 -t 100000 -P 12 -l 30k
>
>Server:
>kernel: 4.0.0-rc4 , when buffer size(-l) >= 30k, no rx_missed_errors
>                     when buffer size(-l) <  30k, confirmed rx_missed_errors
>
>Server:
>kernel: 2.6.32-358 , when buffer size(-l) >= 10k, no rx_missed_errors
>                      when buffer size(-l) <  10k, confirmed rx_missed_errors
>
>Any suggestions?

Monitor the stats while running the test and look for patterns (e.g. rx_missed_errors incrementing along with other stats like tx_xoff, queues etc), you may also want to monitor interrupts/sec and see if it changes with the different flows.

Thanks,
Emil