Testing BQL should happen in openwrt mainline
|Assignee:||Dave Täht||% Done:||
|Category:||Networking||Spent time:||20.00 hours|
|Target version:||1st Public Cerowrt release||Estimated time:||300.00 hours|
BQL's concept is simple and profound enough for others to test in mainline openwrt.
I need to stress that all BQL does at present is move the bufferbloat problem out of the
ethernet drivers and into the packet schedulers, where more packet bursts need to be shredded than
what the default pfifo_fast scheduler can do.
Also: BQL's controller kind of bothers me, but applying hard limits to the amount of bytes in flight
worked well, in testing an earlier version.
- ar71xx working on 3.1.5 - which will contain a fix to RED
- BQL backport to 3.1.5 (I did v3, someone else is doing v4)
- BQL driver support for selected devices (ag71xx) (done, untested)
- Building openwrt with QFQ support in the scheduler (DRR is good too) (patches in cerowrt, currently)
- update to iproute2 to at least the 3.0 level to get QFQ support (also patches in cerowrt, currently)
- Scripts to engage TC and things like QFQ, SFQ, etc on top of line rate and HTB (deBloat.git)
All these patches are floating in loose formation. The ar71xx 3.1 backport arrived this morning...
Useful will be - closely tracking eric dumazet's adaptive red work (which is now in net-next and can also be easily backported), as well as the related iproute2 changes -
most useful of all, for wireless, will be the upcoming 'time in queue' set of ideas, which affect the kernel and iproute as well.
That said, I think all the ideas are promising enough to actually do inside of openwrt, with the help of everyone expending enthusiasm on it all right now, and we will get better and more interesting results, faster.
Doing further work on these exciting bits merely inside of cerowrt would be self limiting.
|related to Cerowrt - Feature #321: ar71xx architecture needs to move into the kernel mainline||New||12/08/2011|