Predicting and Bypassing End-to-End Internet Service Degradations

In this paper the author tried to analyze the possible cause of degradation and how the network might adopt to this degradation, then construct methods to predict these events, and finally, use this predictor to help improve the performance of the gateway.

They define the degradation based on the round-trip-time collected. They made several brave assumptions about the network to simplify their model. For the baseline to define degradation, they use the median or the smallest RTT. Although they argue that median is a fair measurement, their supporting evidence seemed to be weak. They should have provide some longer period statistic like the period and the target of their measurement instead of measurement appeared on the text (90% of the observed difference between median and smallest^Å) since the baseline of defining degradation is the foundation of this paper. Also the capability of the gateways may vary, but they did not give any discussing about the relation between the ability of the gateway verses their defined congestion.

They also define the measurement the predictor rely on, the recall. One potential problem is that since they observe most degradation lay between level one and level six, that means more accurate definition of the first level might be needed. Their model of the host is to divide hosts with several IP address with the same weight. This might deviate from the true situation a lot.

Their define their predictor into two main categories, the single measurement predictor and fixed-window count. Single window measurement based their predictor on certain data measured some time age, and the fix-window measurement is based on the data collected some period of time before. Both measurement are based on past data, and are thus vulnerable when predicting fast-changing events. These predictors might be satisfiable for a static environment but not a highly dynamic one.

Their predictor models are mostly intuitively straight-forward, the difference between different data is just the difference between how the past data are utilized. This is somewhat disappointing. They are just some combination of different math trying to give past data different importance. These can all be used and have some degree of credibility when the network is stable. They still fail to combine the possible behavior of the network that can be viewed as property and can^Òt be observed in some time. For example, if the flow of the network is some square-like pattern, all the predictor can behave well after certain time after the ^Óphase change^Ô. However, non of them can predict for the time of change (predict the time when the flow change from high to low or vice versa. The predictor will make the same mistake of predicting the wrong phase before they adapt to the current flow. And this might be a very unsatisfactory behavior.

The most interesting predictor of all might be the hidden Marcov Model. However ,their explanation on this predictor seemed particularly weak. This paper would seemed to me they are not doing too much beyond-common knowledge breakthrough, and their measuring method is less than perfect, too.