Hi Dave, > On Mar 5, 2015, at 8:58 AM, Dave Taht <[email protected]> wrote: > > I let the data take me where it may. I (not) always have, but reformed > about 15 years ago. [1] I hope that you and your students also, do > some experiments on the successors to PI and RED and DRR - and also > follow the data where-ever it leads you.
In 2003 we had published a paper on STPI (Self-Tuning PI). The self-tuning design accounted for variations in link capacity, presence of cross traffic (i.e. unregulated UDP flows) and variations in the number of flows being controlled. We also proved the (local, exponential) stability of our self-tuning mechanism. The paper is available at http://dna-pubs.cs.columbia.edu/citation/paperfile/79/GLOBECOM_2003_STAQM.pdf This was an evolution over our first PI design where we introduced the concept of linear controllers for AQM. BTW STPI is not PIE, the differences as I see them are (1) STPI explicitly accounts for cross traffic. (2) STPI tunes the parameters continuously, over the entire range of loss (marking) rates unlike PIE which seems to be doing it for 3 specific values chosen ahead of time. (3) The stability of the adaptive mechanism of STPI was rigorously proven in a linearized setting. (4) The tradeoff between stability and responsiveness for the time constant of adaptation was made explicit. (5) STPI does not, as the author of PIE stated recently "control the offset to the reference level and second moment of the latency independently”. STPI simply controls the latency, I don’t know of any way to control the second moment of any reference signal by a linear controller like PI(E), but then I do not know the details of PIE. It was a very deliberate design choice by us to introduce a linear controller like PI for AQM because of ease of implementation. We could’ve gone the route of optimal and/or non-linear controllers but we didn’t. As a side note, we also designed a self-tuning version of RED that we called STRED (different from ARED) but RED suffers from some fundamental limitations so exploring that won’t be of interest. So I am getting STPI and PI implemented as part of cerowrt and will release the code for anyone to play with/evaluate. More than anything, it is the increased deployment of ECN that has revived an interest in AQM for me. -Vishal -- http://www.cs.columbia.edu/~misra/
_______________________________________________ Bloat mailing list [email protected] https://lists.bufferbloat.net/listinfo/bloat
