Monitoring Scheduling in a Kerrighed cluster
Résumé
Kerrighed is a Single System Image project on top of Linux kernel, thus allowing to create a big SMP or cluster on top of several mainstream computers. One component is the scheduler of the cluster. As users can use any node they want to start their process, the load should be unbalanced shortly. The Kerrighed scheduler decides how to balance this load accross the whole cluster (through migration of process for example). Thanks to this mechanism all nodes will be used at their full capacity instead of having one overloaded and the others idling. Prior to Kerrighed 2.4, the scheduler was hard coded. Kerrighed 2.4 introduced SchedConfig a framework allowing the user to create its own scheduler from userspace. We implemented a scheduler, and needed to create new tools to grab data needed to decide wether this scheduler (in its current implementation) is efficient or not. Mainly, the idea is to track migrations that occurs inside the cluster and to log them. This allow to have a trace of the events, and thus to be able to measure some yet-to-be-defined quality measure of the scheduling. We will present how we currently handle migration tracking, and the tools we developped to present it in a viewable manner to the user.