A Graphical Modeling Viewpoint on Queueing Networks

Probabilistic models of the performance of computer systems have long been used to predict future performance. What has not been recognized, however, is that performance models can also be used to diagnose past performance problems. In this paper, we analyze queueing networks from the probabilistic modeling perspective, applying inference methods from graphical models that allow answering diagnostic questions from incomplete data. In particular, we present a slice sampler for networks of G/G/K queues. As an application of this technique, we localize performance problems in distributed systems from incomplete system trace data. On both synthetic networks and a benchmark distributed Web application, we identify bottlenecks with 25% of the overhead of full instrumentation.