Monitoring and Modeling Performance of Communications in Computational Grids

For efficient use of a computational grid, which includes machines located in a number of sites, we have to be able to estimate data delivery times between the machines. For dynamic distributed grids it is unrealistic to know exact parameters of the communication hardware and the current communication traffic, hence we should rely on a performance model of the network to estimate data delivery times. We build an empirical model based on observation of message delivery times with various message sizes and time scales. Our experiments show presence of multiple bands in the histogram of the logarithm of message delivery times. The histograms represent a multiband model reflecting multiple paths and probabilities the messages can travel between the grid machines.