The byte mode being turned off is a pure implementation issue with Linux (when the drop probability is very high, there are some rounding error issues). Vijay can probably articulate it better. The algorithm still has the byte mode operation. I think we can turn it back on for your test where there is no extremely high drop rate and we are working on a fix, which will be available soon.
Thanks, Rong From: Dave Taht <[email protected]<mailto:[email protected]>> Date: Thursday, August 15, 2013 2:41 AM To: Naeem Khademi <[email protected]<mailto:[email protected]>> Cc: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>>, "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>>, Vijay Subramanian <[email protected]<mailto:[email protected]>>, bloat <[email protected]<mailto:[email protected]>> Subject: Re: [aqm] [iccrg] ECN support for PIE AQM in Linux The pie patches that I was using until recently are here: http://snapon.lab.bufferbloat.net/~cero2/deb/patches/ Stephen hemminger and I cleaned up the original pie drop according to KernelStyle in multiple ways. There is a rounding error problem in setting the target value (we'd recomended they switch to ktime rather than jiffies) that I either added or was in there in the first place.... and we fixed a couple generic bugs... In addition to adding ecn support, we also added congestion notification to (a la the linux red and fq_codel algorithms) to provide lossless feedback into the host tcp stack. (this is, incidentally, 1 line of code) This optimization made it possible to do comparisons on host <-> host tests between these qdiscs, with no router in-between. Syntactically the iproute2 patches could use identical syntax for time based values as the codel iproute2 patches. Haven't got around to that. (not my job, really, to improve cisco's code) Now, I only now - today! - just noticed that the new pie code at cisco's site seems *substantially* different from the older pie code the above was based on, in that it no longer has an optimization for not dropping smaller packets. (bytemode) This is going to invalidate most previous benchmarks of pie that I know of, including all of mine, and most, if not all, of cablelabs'. Sigh. I will try to refresh my patches and update the lab over the next week or so. On Wed, Aug 14, 2013 at 2:17 PM, Naeem Khademi <[email protected]<mailto:[email protected]>> wrote: Hi Vijay This code implements basic ECN capability for PIE meaning that it would CE-mark packets that would be early-dropped otherwise. I have sanity-checked the code (e.g. successful ECN marking and receiver/sender reaction). We're now in process of conducting larger set of experiments with PIE w/wo ECN that will eventually be published as we move forward. Cheers, Naeem On Wed, Aug 14, 2013 at 8:27 PM, Vijay Subramanian <[email protected]<mailto:[email protected]>> wrote: > Naeem, > > Thanks for testing and the contribution! We have ECN support in our current > code-under-development thanks to earlier help from Dave Taht but I will > merge your changes as appropriate and > submit it back for review. > > I will take a look at your code soon. Do you also have any test results > which you can share? > > Thanks again, > Vijay > > > On 14 August 2013 07:05, Naeem Khademi > <[email protected]<mailto:[email protected]>> wrote: >> >> Hi >> >> The already-existing PIE Linux code maintained at >> ftp://ftpeng.cisco.com/pie/ does not support ECN. Studying the interaction >> of ECN and AQM algos is probably of interest of some of us. I have added ECN >> support to PIE Linux code and have given it some basic sanity tests. >> >> The modified code is available at >> http://heim.ifi.uio.no/naeemk/research/PIE/ >> >> I would appreciate any feedback. >> >> Regards, >> Naeem >> >> >> >> _______________________________________________ >> iccrg mailing list >> [email protected]<mailto:[email protected]> >> https://www.irtf.org/mailman/listinfo/iccrg >> > _______________________________________________ iccrg mailing list [email protected]<mailto:[email protected]> https://www.irtf.org/mailman/listinfo/iccrg -- Dave Täht Fixing bufferbloat with cerowrt: http://www.teklibre.com/cerowrt/subscribe.html
_______________________________________________ aqm mailing list [email protected] https://www.ietf.org/mailman/listinfo/aqm
