Greetings!
During migration from kernel 3.14 to 4.19, we noticed a regression on
the network performance. Under the exact same circumstances, the
standard deviation of the latency is more than double than before on the
Realtek RTL8111/8168B (10ec:8168) using the r8169 driver.
Kernel 3.14:
# netperf -v 2 -P 0 -H <netserver-IP>,4 -I 99,5 -t omni -l 1 -- -O
STDDEV_LATENCY -m 64K -d Send
313.37
Kernel 4.19:
# netperf -v 2 -P 0 -H <netserver-IP>,4 -I 99,5 -t omni -l 1 -- -O
STDDEV_LATENCY -m 64K -d Send
632.96
In contrast, we noticed small improvements in performance with other
non-Realtek network cards (igb, tg3). Which suggested a possible driver
related bug.
However after bisecting the code, I ended up with the following patch,
which was introduced in kernel 4.17 and modifies net/ipv4:
commit 0a6b2a1dc2a2105f178255fe495eb914b09cb37a
Author: Eric Dumazet <eduma...@google.com>
Date: Mon Feb 19 11:56:47 2018 -0800
tcp: switch to GSO being always on
Could you please help me to clarify, should GSO be always on on my
device? Or does it just affect TCP? According to ethtool it is always
off, "ethtool -K eth0 gso on" has no effect, unless I switch SG on.
# ethtool -k eth0
Offload parameters for eth0:
Cannot get device udp large send offload settings: Operation not
supported
rx-checksumming: on
tx-checksumming: off
scatter-gather: off
tcp-segmentation-offload: off
udp-fragmentation-offload: off
generic-segmentation-offload: off
generic-receive-offload: on
large-receive-offload: off
I validated that reverting "tcp: switch to GSO being always on"
successfully brings back the better performance for the r8169 driver.
I'm sure that reverting that commit is not the optimal solution, so I
would like to kindly ask for help to shed some light in this issue.
Best regards,
Juliana.
