pps (a tool developed by P. Saab and previously used and improved by K. Kennaway) blasting UDP packets in parallel threads is testing UDP send performance only. We run it to a local disc(4) (discard interface) for these tests, no even sending packets on the wire. We ran 3 reboot iterations, with each iteration having 1 warm up run and 10 recorded pps runs of 120 seconds.
On IPv6 UDP checksums are mandatory, so we did not turn them off for comparisons.
As we can see, turning off checksumming on IPv4 does not make a difference for this test. However IPv6 UDP is now ahead of IPv4 UDP on this hardware. On other machines tested the gap is not as significant but in no case there had been an IPv6 regression.
x hydra1/test_pps_udp4_local_discard.stat
+ hydra1/test_pps_udp4_local_discard_nocksum.stat
* hydra1/test_pps_udp6_local_discard.stat
+--------------------------------------------------------------------------+
| * |
| * |
| * |
| * +x |
| * * +x x|
| * * +* x|
| * * +* x|
| * * +* *|
| ** * +* *|
| ** * **+*|
| ** * **+*|
|* ** * + ****|
|* ** ** * ****|
| |A_||
| |_A| |
| |_MA_| |
+--------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 30 288647.14 297221.37 294965.73 294963.19 1826.9506
+ 30 288820.84 296645.88 294073.88 294086.6 2073.5719
No difference proven at 95.0% confidence
* 30 191089 202035.53 194729.67 196752.63 3732.0367
Difference at 95.0% confidence
-98210.6 +/- 1518.79
-33.2959% +/- 0.514908%
(Student's t, pooled s = 2938.18)