Toke Høiland-Jørgensen <t...@toke.dk> writes: > On 10 August 2016 21:35:40 CEST, Dave Taht <dave.t...@gmail.com> wrote: >>Wow, that *is* weird. It is good to see the tcp window changing on >>this set of data (it wasn't before), and CWRs, but... hmmm... SCIENCE. >> >>Enabling ecn on both sides will rule out some potential bugs. > > Yeah, couldn't get ecn to work on the host I was using as the other endpoint > on > that test. Will try with another box that's not on quite as ancient a kernel. > Was also planning to disable codel (by setting a very high target) to try to > narrow down the problem.
OK, digging some more on this: I am seeing *no* drops by CoDel, and no backlog in the mac80211 softq layer either (or at most one or two packets when polling with a 1 sec interval). This is with one as well as with two flows. On my x86 testbed I see backlog building and packets getting dropped by CoDel - and can't reproduce the performance hit. So I'm wondering what the difference is. I can think of: - Another bottleneck somewhere in the system limiting things on the wndr3800. - A locking issue in the FQ mechanism preventing packets from being queued properly. - Something related to the wndr only having a single CPU. -Toke _______________________________________________ Cerowrt-devel mailing list Cerowrt-devel@lists.bufferbloat.net https://lists.bufferbloat.net/listinfo/cerowrt-devel