On 08/01/15 01:20, Fred Baker (fred) wrote:
>> Yes. I think something like this would be good. The buffer bloat >> example is probably an extreme case of things not having AQM at all >> and interacting badly. It would maybe be worth mentioning that any >> AQM mechanism has also got to work in series with boxes that don't >> have any active AQM - just tail drop. Ultimately, I would say this >> is just a matter of control engineering principles: You are >> potentially making a network in which various control algorithms >> are implemented on different legs/nodes and the combination of >> transfer functions could possibly be unstable. Has anybody applied >> any of the raft of control theoretic methods to these algorithms? >> I have no idea! > Well, PIE basically came out of control theory (a basic equation that > describes a phase-locked loop), and I believe that Van and Kathy will > say something similar about Codel. But that’s not a question for this > paper, it’s a question for the various algorithmic papers. Elwyn, I would like to expand on a comment that Fred made above. He is absolutely correct, about PIE coming out of control theory. PIE is based on the PI AQM algorithm proposed by us back in 2001 (http://dna-pubs.cs.columbia.edu/citation/paperfile/23/MisraInfocom01-AQM-Controller.pdf <http://dna-pubs.cs.columbia.edu/citation/paperfile/23/MisraInfocom01-AQM-Controller.pdf>), where we showed the benefits of using a PI controller to regulate queue lengths and delays through a formal control theoretic analysis. The PI algorithm itself came from a fluid model of TCP+AQM that we developed (http://dna-pubs.cs.columbia.edu/citation/paperfile/27/sigcomm2000.pdf <http://dna-pubs.cs.columbia.edu/citation/paperfile/27/sigcomm2000.pdf>) and the model is referenced in the IETF draft. When we designed PI, we took great care in designing an algorithm that was amenable to a very simple implementation at line speeds. We could have gone in the direction of sophisticated (non-linear/optimal) control algorithms but we instead kept simplicity of implementation front and center and it is great to see that the standards community has come to appreciate the work. The PIE update algorithm is very similar to the original PI algorithm (we controlled queue length to a reference value whereas PIE controls delay by dividing the queue length by measured link capacity) - Rong Pan and the PIE team have done great work in adapting the originial algorithm for implementation in production environments and all our original analyses carries through to PIE. I am happy to provide references to more control theoretic analyses (including ours) that look at the behavior and stability of the AQM algorithms. One thing to note though that hasn't changed all these years is if you look at Section VII.A of our PI paper linked above, the full benefits of AQM are realized in conjunction with ECN. On a bottlenecked link, if you reduce delay (by controlling it via a mechanims like PI(E) or CoDel), unless you have ECN implemented you will end up increasing loss rates which may not be a good thing. Lastly, I had contributed pi.cc <http://pi.cc/> to ns2 back in 2001 - a version seems to have survived all these years if anyone is interested in playing with it: https://github.com/jridgewell/ns2/blob/master/queue/pi.cc <https://github.com/jridgewell/ns2/blob/master/queue/pi.cc> Thanks, -Vishal -- http://www.cs.columbia.edu/~misra/
_______________________________________________ aqm mailing list [email protected] https://www.ietf.org/mailman/listinfo/aqm
