I just pushed out those skb memory use reduction hacks. It would be good if more were to try them, in std openwrt, and also a patch for pfifo_fast produced, for openwrt's general use.
https://github.com/dtaht/Cerowrt-3.3/commit/d31d3faa0b7cf7254e6b657d85e8e5168ab55866 in some thread around here or the codel list was eric's proposed patch reducing ath9k's skb pre-allocation size. I agree with eric in that the current allocation method doesn't make any sense... (and see comments in above commit) Patch 53 was highly experimental and should be ignored. I have something better now, but it ain't ready. On Thu, Sep 6, 2012 at 9:20 AM, Dave Taht <[email protected]> wrote: > The ubnt builds are a bit out of date (part of my test deployment of > fq_codel at a campground here - the yurtlab! > http://snapon.lab.bufferbloat.net/~d/lupin/yurtlab.jpg which is > blessedly free > of competing wifi signals) > > They are also a little specialized. Among other things, they have my > ssh public key in them, > and are keyed to local dns, they rely on the babel protocol to route > and mesh together, although they do come up presently on a fixed ip > address that needs to be changed, instead of relying on AHCP for their > IPs and dns. > > All of quagga (including OSPF) IS built, it's just that babel is the default. > > In working with this build and test deployment I exposed some problems > with memory usage in codel and the ath9k driver. Quagga will get > OOM-killed under heavy loads, such as RTP and UDP flooding across > multiple diffserv values, on the 2HP boxes in AP mode. The p2p > backbone (nanostation M5s) has never crashed on me. > > This points long term towards developing a "mfq_codel", which would > have one qdisc and *one* set of buffers that controls all four > hardware queues. However, try as I might, I haven't wrapped my head > around how to merge the code for sch_mq and sch_fq_codel, in a way > that would work. (?) > > Aside from that it's working rather well, and the hacks I have in > place now to reduce skb size seem to help a lot. > > Now that the memory problems are mostly beaten I will be doing a > respin and re-deployment towards the end of the month. If I can build > something a little more generic and easy to deploy on ubnt for others, > I will do so, getting to where all this code runs well in 32MB is > rather important to me. > > Let me know. > > On Thu, Sep 6, 2012 at 4:17 AM, Andy S <[email protected]> wrote: >> I see that there are UBNT builds available for specific deployment >> scenarios; we use Ubiquiti Bullet M5's on some parts of our (city-wide) >> 5.8Ghz wireless network, but they don't support OSPF, so if any of the >> builds of Cerowrt support Bullet M5 and OSPF (I see the Quagga-ospf modules >> available in the main devel builds) - we would be interested to hear/test! >> >> Cheers, >> >> A >> -- >> http://www.bristolwireless.net/ | 0117 325 0067 | >> sip:[email protected] >> Bristol Wireless >> Windmill Hill City Farm >> Philip Street >> Bedminster >> Bristol BS3 4EA >> >> >> _______________________________________________ >> Cerowrt-devel mailing list >> [email protected] >> https://lists.bufferbloat.net/listinfo/cerowrt-devel >> > > > > -- > Dave Täht > http://www.bufferbloat.net/projects/cerowrt/wiki - "3.3.8-17 is out > with fq_codel!" -- Dave Täht http://www.bufferbloat.net/projects/cerowrt/wiki - "3.3.8-17 is out with fq_codel!" _______________________________________________ Cerowrt-devel mailing list [email protected] https://lists.bufferbloat.net/listinfo/cerowrt-devel
