A High-Performance Sensor for Cluster Monitoring and Adaptation

As Beowulf clusters have grown in size and complexity, the task of monitoring the performance, status, and health of such clusters has become increasingly more difficult but also more important. Consequently, tools such as Ganglia and Supermon have emerged in recent years to provide the robust support needed for scalable cluster monitoring. However, the scalability comes at the expense of accuracy in that the tools only obtain data samples through an entry in the /proc filesystem and only at the granularity of a kernel tick, i.e., 10 milliseconds. As an alternative to using /proc as a sensor for Ganglia and Supermon, we propose a dynamic, high-fidelity, event-based sensor called MAGNET (Monitoring Apparatus for General kerNelEvent Tracing). Unlike our previous incarnation of MAGNET, this incarnation allows for the dynamic insertion and deletion of instrumentation points and improves performance by approximately 100% over our previously low-overhead MAGNET and approximately 25% over the Linux Trace Toolkit (LTT) while providing superior functionality and robustness over LTT. Furthermore, our latest MAGNET is flexible enough to morph itself into other tools such as tcpdump and yet still high performance enough to perform over 250% better than tcpdump. It can also be used as a diagnostic (or debugging) tool, a performance-tuning tool, or a reflective tool to enable self-adapting applications in clusters or grids.