Bugzilla – Bug 2593
Lower wifi performance since commit 12345:a94d790ef6e5 (Bug 2369)
Last modified: 2016-12-22 17:16:30 UTC
I notice performance reduction since commit 12345:a94d790ef6e5 for bug 2369. For example, single 802.11a station with null distance: Before commit: 29.8914 Mbit/s After commit: 29.5717 Mbit/s Bianchi expected throughput: 29.9 Mbit/s I guess this throughput reduction is even more significant at higher bitrates (802.11n/ac). Could Tom and/or Ali investigate?
Patch urgently wanted since I cannot check whether no other regression are introduced when looking at wifi performance. Please let me know if any support is needed, but since this comes from a commit from Tom and Ali, I guess you could quickly find what is wrong. In case you come to the conclusion we were overestimating throughput before those changes, then I can quickly provide a comparison between Bianchi model and our previous and new results. But my feeling is that we were much closer to the Bianchi curve before that commit.
(In reply to sebastien.deronne from comment #0) > I notice performance reduction since commit 12345:a94d790ef6e5 for bug 2369. > > For example, single 802.11a station with null distance: > Before commit: 29.8914 Mbit/s > After commit: 29.5717 Mbit/s > Bianchi expected throughput: 29.9 Mbit/s > > I guess this throughput reduction is even more significant at higher > bitrates (802.11n/ac). > > Could Tom and/or Ali investigate? Can you provide a test script and clarify what you mean by Bianchi expected throughput for single station with null distance? I understand Bianchi to apply to an environment with multiple stations, at least the IEEE JSAC paper from 2000. We should also check that the current model is consistent with Bianchi's assumptions since we introduced a behavior (busy during AIFS) that is needed for broadcast relaying; a scenario that is not covered by Bianchi AFAIK. It would be helpful to have a clear Bianchi-based example in our codebase, including EDCA, with backoff statistics, and then to put the output into regression testing. We had a poster on this topic back in WNS3 2012 but no contributed code on it has made it to ns-3-dev.
(In reply to Tom Henderson from comment #2) > (In reply to sebastien.deronne from comment #0) > > I notice performance reduction since commit 12345:a94d790ef6e5 for bug 2369. > > > > For example, single 802.11a station with null distance: > > Before commit: 29.8914 Mbit/s > > After commit: 29.5717 Mbit/s > > Bianchi expected throughput: 29.9 Mbit/s > > > > I guess this throughput reduction is even more significant at higher > > bitrates (802.11n/ac). > > > > Could Tom and/or Ali investigate? > > > Can you provide a test script and clarify what you mean by Bianchi expected > throughput for single station with null distance? I understand Bianchi to > apply to an environment with multiple stations, at least the IEEE JSAC paper > from 2000. Yes I expressed myself badly :-) I actually have a bianchi matlab script that computes the expected throughput for a given number of parameters. If I set it to N=1, it does not really use the Bianchi model, so please forget about that Bianchi word when I talk about one station. > > We should also check that the current model is consistent with Bianchi's > assumptions since we introduced a behavior (busy during AIFS) that is needed > for broadcast relaying; a scenario that is not covered by Bianchi AFAIK. Ok, then it might be expected that we diverge from the model. Note that I notice that we were slightly overestimating the expected throughput for large number of stations, so maybe we will still be close for a large number of stations. I will retry and if acceptable, I will reject this bug. > > It would be helpful to have a clear Bianchi-based example in our codebase, > including EDCA, with backoff statistics, and then to put the output into > regression testing. We had a poster on this topic back in WNS3 2012 but no > contributed code on it has made it to ns-3-dev. I plan to deliver it someday, I have a draft ready but I need to do some more work. For the moment it plots the throughput versus number of stations and I manually compare against Matlab results. Might be nice it we would have something totally automated and if we would provide more statistics. Do you have any support about that WNS3 contribution?
This bug can be rejected, performance are still fitting well the one computed with the Bianchi model when I consider a large number of stations. Furthermore the difference is not really significant.