netperf-2.5 TCP_STREAM test was run over 3 reboot iterations, with each iteration having 1 warm up run and 10 recorded netperf runs.
As we can see, adding support for (pretended) "checksum offloading" for IPv6 made a huge difference compared to the previous results, where the medians were basically equal.
x hydra1/test_netperf_tcp6_default.stat + hydra1/test_netperf_tcp6_nooffload.stat +--------------------------------------------------------------------------+ | + | | ++ | | ++ | | +++ + + x x x | | ++++ ++++ xxx x x x x | |++++++++++++ xxxxxxx xxxx x xxxxx x xx| | |_______MA________| | | |_MA__| | +--------------------------------------------------------------------------+ N Min Max Median Avg Stddev x 30 6382.25 7656.47 6855.85 6913.7367 379.66968 + 30 4474.78 4960.58 4632.14 4678.9613 135.61332 Difference at 95.0% confidence -2234.78 +/- 147.361 -32.3237% +/- 2.13143% (Student's t, pooled s = 285.079)
Given IPv6 "offload support" has been implemented, we added the 4th value now also distinguishing IPv6 with and without offload support..
As expected IPv4 and IPv6 without offload differ to with offloading. Zooming in to our surprise we find that IPv6 with offload support is doing slightly better than IPv4 with offload support. However without offload support IPv4 is still doing a lot better.
x hydra1/test_netperf_tcp6_nooffload.stat + hydra1/test_netperf_tcp4_nooffload.stat * hydra1/test_netperf_tcp6_default.stat % hydra1/test_netperf_tcp4_default.stat +--------------------------------------------------------------------------+ | %% | | xxx xx ++ ++ %%O* *% * | | xxxx xxx x +++ ++ + %%%OO * **% *O*% | |xxxxxxxxxxx + ++++++++++ + %%OOOOOO% O**O % * *O*O * *O %| | |_MA__| | | |_AM_| | | |______MA_______| | | |____M___A________| | +--------------------------------------------------------------------------+ N Min Max Median Avg Stddev x 30 4474.78 4960.58 4632.14 4678.9613 135.61332 + 30 5527.42 6112.51 5834.32 5804.6183 129.32283 Difference at 95.0% confidence 1125.66 +/- 68.4939 24.0578% +/- 1.46387% (Student's t, pooled s = 132.505) * 30 6382.25 7656.47 6855.85 6913.7367 379.66968 Difference at 95.0% confidence 2234.78 +/- 147.361 47.7622% +/- 3.14945% (Student's t, pooled s = 285.079) % 30 6306.3 7892.74 6527.82 6729.2337 424.53354 Difference at 95.0% confidence 2050.27 +/- 162.898 43.819% +/- 3.48149% (Student's t, pooled s = 315.135)