2008-09-11 Readings - Fair Queueing

ANALYSIS AND SIMULATION OF A FAIR QUEUEING ALGORITHM

Alan Demers, Srinivasan Keshav, Scott Shenker


CORE-STATELESS FAIR QUEUEING: ACHIEVING APPROXIMATELY FAIR BANDWIDTH ALLOCATION IN HIGH SPEED NETWORKS

Ion Stoica, Scott Shenker, Hui Zhang


I'll talk about both papers together. I'm not sure what to say about these papers. I had something that was at the same time a very strong reaction and no reaction at all. This confusion is due to several reasons. First, I'm not sure about the importance of the topic. Is it the case that FIFO is still by far the predominant queueing mechanism, and it has more or less worked so far? Second, I'm not sure if the need for fair queueing has been sufficiently motivated. The unresponsive media applications argument seems to move towards vague QoS regions ... The uncooperative flows motivation seems to be the only convincing one ... I'll look at readings from Vern Paxson's web security class to see how they deal with that. Third, both papers have so many different results from so many different angles that the only bottom line I could extract is fair queueing - Ok it's better than FIFO, but how much better? Depends a great deal on topology/assumptions/traffic. And all the other fair queueing mechanisms in their comparative studies have their own trade offs as well. So it's not clear how much fair queueing would buy you at what cost. Forth, the CSFQ and presumable other flow-based FQ mechanisms rely on a way to label flows, which rely on a way to define flows. Both papers have stated the difficulty of a good definition for flows, and the lack of a way to identify and label flows. So shouldn't flow definition/identification/labeling be a pre-requisite study? Lastly, I can think of a whole host of reasons that ISPs would refuse to adopt fair queueing for non-technology concerns. So is the whole topic a field of research that did not get used, get developed with insufficient motivation, has unclear outcomes, depend on non-existent technology, and wouldn't get adopted in real life anyway even if it works? I'm hesitant to think Ion and Scott would do research with very little to gain ...

Hence, very much look forward to class discussion to clear my thoughts.