Using Search Methods for Selecting and Combining Software Sensors to Improve Fault Detection in Autonomic Systems

Fault-detection approaches in autonomic systems typically rely on runtime software sensors to compute metrics for CPU utilization, memory usage, network throughput, and so on. One detection approach uses data collected by the runtime sensors to construct a convex-hull geometric object whose interior represents the normal execution of the monitored application. The approach detects faults by classifying the current application state as being either inside or outside of the convex hull. However, due to the computational complexity of creating a convex hull in multi-dimensional space, the convex-hull approach is limited to a few metrics. Therefore, not all sensors can be used to detect faults and so some must be dropped or combined with others. This paper compares the effectiveness of genetic-programming, genetic-algorithm, and random-search approaches in solving the problem of selecting sensors and combining them into metrics. These techniques are used to find 8 metrics that are derived from a set of 21 available sensors. The metrics are used to detect faults during the execution of a Java-based HTTP web server. The results of the search techniques are compared to two hand-crafted solutions specified by experts.