On Wed, May 30, 2018 at 6:36 AM Eric Dumazet [off-list ref] wrote:

Here are the good ones, using latest David Miller net tree. ( plus
https://patchwork.ozlabs.org/patch/922528/  but that should not matter

here)

llpaa23:/export/hda3/google/edumazet# ./netperf -H 2607:f8b0:8099:e18:: -t
UDP_STREAM
MIGRATED UDP STREAM TEST from ::0 (::) port 0 AF_INET6 to
2607:f8b0:8099:e18:: () port 0 AF_INET6
Socket  Message  Elapsed      Messages
Size    Size     Time         Okay Errors   Throughput
bytes   bytes    secs            #      #   10^6bits/sec

212992   65507   10.00      216236      0    11331.89
212992           10.00      215068           11270.68

There are few drops because of the too small
/proc/sys/net/core/rmem_default  ( 212992 as seen in netperf output) for
these kind of stress.
( each 64KB datagram actually consumes half the budget ...)

Once rmem_default is set to 1,000,000 and  mtu set back to 1500 (instead of
5102 on my testbed)
results are indeed better.

lpaa23:/export/hda3/google/edumazet# ./netperf -H 2607:f8b0:8099:e18:: -t
UDP_STREAM -l 10
MIGRATED UDP STREAM TEST from ::0 (::) port 0 AF_INET6 to
2607:f8b0:8099:e18:: () port 0 AF_INET6
Socket  Message  Elapsed      Messages
Size    Size     Time         Okay Errors   Throughput
bytes   bytes    secs            #      #   10^6bits/sec

212992   65507   10.00      231457      0    12129.56
1000000           10.00      231457           12129.56