dcbsdcon 2009
brooks davis
isolating cluster jobs for performance
and predictability
At The Aerospace Corporation, we run a large FreeBSD based computing cluster to support engineering applications. These applications come in all shapes, sizes, and qualities of implementation. To support them and our diverse userbase we have been searching for ways to isolate jobs from one another in ways that are more effective than Unix time sharing and more fine grained than allocating whole nodes to jobs. In this paper we discuss the problem space and our efforts so far. These efforts include implementation of partial file systems vitalization and CPU isolation using CPU sets.
return to speakers and talks