Comment # 1 on bug 1203603 from Jiri Wiesner

It should be noted that moving both the client and the server iperf3 process to
NUMA node 0 causes a performance hit, which is not as pronounced as in the TCP
test:
> simba3:~/:[1]# taskset 0x0000ffff /root/jwiesner/iperf/src/iperf3 -u -c 10.100.128.66 -P1 -f m -b 0 -t 180 -i 5 -O 2
> Connecting to host 10.100.128.66, port 5201
> [  5] local 10.100.128.68 port 42650 connected to 10.100.128.66 port 5201
> [ ID] Interval           Transfer     Bitrate         Total Datagrams
> [  5]   0.00-5.00   sec  4.15 GBytes  7121 Mbits/sec  3829980
> [  5]   5.00-10.00  sec  4.14 GBytes  7114 Mbits/sec  3070640
> [  5]  10.00-15.00  sec  4.15 GBytes  7129 Mbits/sec  3076920

The MTU of all devices was set to 1500 bytes. This is nstat output for a 10
second period on simba2 for the UDP test:
> IpInReceives                    7951006            0.0
> IpInDelivers                    7950996            0.0
> IpOutRequests                   2                  0.0
> TcpInSegs                       2                  0.0
> TcpOutSegs                      2                  0.0
> UdpInDatagrams                  7078508            0.0
> UdpInErrors                     872514             0.0
> UdpRcvbufErrors                 872514             0.0
> Ip6InReceives                   16                 0.0
> Ip6InMcastPkts                  16                 0.0
> Ip6InOctets                     1422               0.0
> Ip6InMcastOctets                1422               0.0
> Ip6InNoECTPkts                  16                 0.0
> TcpExtTCPHPAcks                 2                  0.0
> TcpExtTCPAutoCorking            1                  0.0
> TcpExtTCPOrigDataSent           2                  0.0
> TcpExtTCPDelivered              2                  0.0
> IpExtInOctets                   11735864062        0.0
> IpExtOutOctets                  1432               0.0
> IpExtInNoECTPkts                7951140            0.0
The same output for the TCP test:
> IpInReceives                    1812332            0.0
> IpInDelivers                    1812320            0.0
> IpOutRequests                   496318             0.0
> TcpInSegs                       1812324            0.0
> TcpOutSegs                      496319             0.0
> Ip6InReceives                   13                 0.0
> Ip6InMcastPkts                  13                 0.0
> Ip6InOctets                     1115               0.0
> Ip6InMcastOctets                1115               0.0
> Ip6InNoECTPkts                  13                 0.0
> TcpExtTCPHPHits                 1812243            0.0
> TcpExtTCPPureAcks               7                  0.0
> TcpExtTCPHPAcks                 4                  0.0
> TcpExtTCPAutoCorking            1                  0.0
> TcpExtTCPOrigDataSent           10                 0.0
> TcpExtTCPDelivered              11                 0.0
> IpExtInOctets                   116209437656       0.0
> IpExtOutOctets                  25809812           0.0
> IpExtInNoECTPkts                80194342           0.0
The IP layer in the UDP test processes MTU-sized packets:
11735864062 (IpExtInOctets) / 7950996 (IpInDelivers) = 1476 bytes per packet
whereas IP layer in the TCP test processes much larger packets approaching the
maximum size of an IP packet:
116209437656 (IpExtInOctets) / 1812320 (IpInDelivers) = 64121 bytes per packet