Towards a Bayesian Statistical Model for the Classification of the Causes of Data Loss

Given the critical nature of communications in computational Grids it is important to develop efficient, intelligent, and adaptive communication mechanisms. An important milestone on this path is the development of classification mechanisms that can distinguish between the various causes of data loss in cluster and Grid environments. The idea is to use the classification mechanism to determine if data loss is caused by contention within the network or if the cause lies outside of the network domain. If it is outside of the network domain, then it is not necessary to trigger aggressive congestion-control mechanisms. Thus the goal is to operate the data transfer at the highest possible rate by only backing off aggressively when the data loss is classified as being network related. In this paper, we investigate one promising approach to developing such classification mechanisms based on the analysis of the patterns of packet loss and the application of Bayesian statistics.