Thank you Toke (and Michael) for the pointers. Just based on the results of the IETF 91 slides, it would seem that disabling CoDel in fq_codel helps in a variety of cases (most evidently in the fairness plots), and only causes some increase in VoIP delay at relatively low link rates (the fq_codel and fq_nocodel curves do overlap in the 100/100 Mbps case of slide 34).
CoDel's effort to limit the queuing delay appears well motivated when there is only one queue (e.g., to shield VoIP from the delay induced by bulk-transfer or video traffic), but when multiple queues are available the net gain of the effort is unclear. I suspect the same may happen with the flow-queue version of PIE. Is this just because the set of tests in the slides does not include one that clearly exposes the superiority of fq_codel over fq_nocodel? And what would that test look like? Andrea -----Original Message----- From: [email protected] [mailto:[email protected]] Sent: Monday, April 13, 2015 11:54 AM To: Francini, Andrea (Andrea) Cc: [email protected] Subject: Re: [aqm] References on AQM test results and fq_nocodel Hi Andrea > 1. Besides the results in Toke’s ICCRG presentation at IETF 91 in > November 2014 > (http://www.ietf.org/proceedings/91/slides/slides-91-iccrg-4.pdf), > where can I find other comprehensive comparisons of AQM > implementations for Linux (or any other non-simulated system that > handles real packets)? The data underlying my experiments are available from here: http://www.cs.kau.se/tohojo/modern-aqms/ -- that page also has links to the test scripts used to run the experiments. The paper is still under submission, so can't link that, sadly, but I can share a copy with you privately if you are interested. > 2. In Toke’s slides, the fq_nocodel scheme appears to be never worse > than fq_codel (or any other AQM in those tests), except for the VoIP > delay plot of slide 12. Besides the Linux code, is there anywhere a > description of the fq_nocodel scheme? Well, fq_nocodel is basically the term I use for fq_codel configured so as to disable the CoDel algorithm (i.e. the target and/or interval is set to 100 seconds so the algorithm never triggers). As such, it doesn't have a description outside of my paper, but you can read the fq_codel draft and ignore everything it says about CoDel, I suppose :) Also, note that configuring fq_codel in this way in Linux is not optimal in terms of CPU use: it deliberate triggers the overflow drop mechanism, which does a linear search through all configured queues, eating up 4k of cache on every drop. If your test router is sufficiently CPU-over-provisioned (a modern x86 counts as this for <=1Gbit speeds) this doesn't affect performance, but I wouldn't recommend it on a tiny MIPS processor, for instance. (Oh, and this limitation is specific to the current implementation, and could easily be fixed if anyone wanted to; the overflow mechanism was meant as a backup, and so wasn't optimised like, say, the SFQ was). -Toke _______________________________________________ aqm mailing list [email protected] https://www.ietf.org/mailman/listinfo/aqm
