Towards a bayesian statistical model for the classification of the causes of data loss
HPCC'05 Proceedings of the First international conference on High Performance Computing and Communications
| Hi-index | 0.00 |
In the context of TCP, several researchers have proposed heuristics to detect or predict congestion in the network. In this paper, the term , refers to such heuristics. Past proposals require TCP sender to reduce its window size when congestion is detected or predicted (otherwise, the heuristic may dictate that the sender window be held constant or increased). The proposed heuristics to detect/predict congestion typically use simple statistics on observed round-trip times and/or observed throughput. The primary objective of this paper is to investigate the ability of the congestion predictors to predict a packet loss. Our measurements indicate that the three congestion predictors studied in this paper are often poor in their ability to predict a packet loss due to congestion. To arrive at this conclusion we measure the frequency with which the predictors predict congestion, and how often they predict congestion just before a packet loss. A study of the variations in measured parameters, as a function of several network parameters (for instance, router queue size) yields several interesting observations as reported in the paper. Although the results presented here are not related to wireless communication, this research was motivated by a desire for an end-to-end mechanism for differentiating between packet losses due to congestion and packet losses due to wireless transmission errors. One technique we considered would use congestion predictors for this purpose. The results presented here suggest that simple congestion predictors will not be effective in differentiating between the two forms of losses.