Latency Prediction Using Aggregate Latency Profiles

A prominent challenge in the support of Wide Area Applications (WAA) involves their unpredictable behavior over a dynamic WAN that results in a considerable variability in access latency (end-toend delay). Latency profiles capture the changing latencies that clients experience when accessing a server and can be utilized as a WAA monitoring and optimization tool. However, in the presence of hundreds of servers and tens of thousands of clients, managing millions of latency profiles cannot scale. In this paper we propose a method for scalable monitoring and performance prediction of WAA using aggregate latency profiles. We utilize Relevance Networks to computes and analyzes comprehensive pair-wise mutual information and correlation for the latency profiles, as well as to build and maintain aggregate latency profiles. We have conducted experiments demonstrating that there is a considerable amount of non-random associations between individual latency profiles. We empirically show that aggregating them improves the overall prediction power of distributed performance monitoring for WAAs.