Abstract

Diagnosis of performance problems in computer systems is a rich application area for machine learning, because data about system performance can be readily obtained. Many such diagnostic questions concern system performance in the face of load, for example, diagnosing performance bottlenecks during past workload spikes, or diagnosing bottlenecks during particularly slow requests. Diagnostic problems are also of practical interest, because much of the cost of managing a system is due to finding and recovering from failures.

Model-based approaches are attractive here, because they can incorporate human knowledge about the system, and do not require labeled failure data. A classical family of models of computer performance is queueing models. Queueing models predict the explosion in system latency under high workload in a way that is often reasonable for real systems.

In this talk, we present a novel graphical modeling viewpoint on queueing models, which allows them to be used for inference about past system behavior and learning from incomplete data. The idea is to measure a small set of arrival and departure times from the system, treating the times that were not measured as missing data. The posterior distribution over missing data and parameters can then be sampled using Markov chain Monte Carlo techniques. Developing a sampler in this case is significantly more challenging than for standard graphical models, because of the complex deterministic dependencies that arise in queueing models. On data from a benchmark Web 2.0 application, we demonstrate the ability to localize performance problems with 25% of the measurement overhead.