On Wed, Aug 28, 2013 at 10:54:42PM +0200, Antonio Quartulli wrote:
Hello list,
This is the *fifth* version of the "improved routing abstraction" patchset. Thanks for the feedback so far! (Message below is the same as v3)
This night I performed another series of tests using two OM2P-HSs meshing over the Ethernet interface. The tests have been performed by using iperf which generated a UDP stream from a node to the other.
After discussing with Marek we decided to use iperf because it would have generated much more CPU load than a "client-to-client" test and would have probably helped in showing the performance loss given by the patchset.
The command used was "iperf -uc 192.168.103.2 -b 85M". Bandwidth higher than 85M was generating packet loss >20% so I decided to do not go beyond that value.
The results are: without abstraciton patchset:
[ 3] 0.0-10.0 sec 98.3 MBytes 82.4 Mbits/sec 0.054 ms 406/70494 (0.58%) [ 3] 0.0-10.0 sec 98.5 MBytes 82.6 Mbits/sec 0.058 ms 726/70999 (1%) [ 3] 0.0-10.0 sec 98.8 MBytes 82.9 Mbits/sec 0.077 ms 422/70931 (0.59%) [ 3] 0.0-10.0 sec 98.7 MBytes 82.8 Mbits/sec 0.078 ms 550/70938 (0.78%) [ 3] 0.0-10.0 sec 98.7 MBytes 82.8 Mbits/sec 0.057 ms 380/70795 (0.54%) [ 3] 0.0-10.0 sec 98.1 MBytes 82.3 Mbits/sec 0.053 ms 516/70499 (0.73%) [ 3] 0.0-10.0 sec 98.6 MBytes 82.7 Mbits/sec 0.088 ms 267/70614 (0.38%) [ 3] 0.0-10.0 sec 98.4 MBytes 82.6 Mbits/sec 0.056 ms 515/70726 (0.73%) [ 3] 0.0-10.0 sec 98.5 MBytes 82.7 Mbits/sec 0.044 ms 526/70817 (0.74%) [ 3] 0.0-10.0 sec 98.3 MBytes 82.5 Mbits/sec 0.044 ms 691/70811 (0.98%)
with abstraction patchset: [ 3] 0.0-10.0 sec 98.0 MBytes 82.2 Mbits/sec 0.113 ms 671/70568 (0.95%) [ 3] 0.0-10.0 sec 98.7 MBytes 82.8 Mbits/sec 0.085 ms 600/70989 (0.85%) [ 3] 0.0-10.0 sec 98.5 MBytes 82.6 Mbits/sec 0.088 ms 496/70725 (0.7%) [ 3] 0.0-10.0 sec 98.5 MBytes 82.6 Mbits/sec 0.063 ms 537/70784 (0.76%) [ 3] 0.0-10.0 sec 98.4 MBytes 82.6 Mbits/sec 0.093 ms 589/70798 (0.83%) [ 3] 0.0-10.0 sec 98.1 MBytes 82.3 Mbits/sec 0.060 ms 761/70742 (1.1%) [ 3] 0.0-10.0 sec 98.0 MBytes 82.2 Mbits/sec 0.098 ms 764/70674 (1.1%) [ 3] 0.0-10.0 sec 98.5 MBytes 82.7 Mbits/sec 0.063 ms 761/71047 (1.1%) [ 3] 0.0-10.0 sec 97.9 MBytes 82.2 Mbits/sec 0.067 ms 953/70821 (1.3%) [ 3] 0.0-10.0 sec 98.0 MBytes 82.2 Mbits/sec 0.070 ms 697/70620 (0.99%)
Given the results I would say that the patchset is not introducing any performance loss.
By the way I am not 100% that the overhead brought by iperf is comparable to the overhead generated by the abstraction code..if not, the consequence could be that iperf is hiding the overhead given by the patchset.
However we can say that the patchset is not introducing any big performance drop and this is a good point! :-)
Cheers,