Scheduling Support for High Performance Clusters

Clusters are becoming a predominant and cost-effective platform for high performance parallel computing. These systems are gaining acceptance not just for scientific applications, which have traditionally needed supercomputers, but are also being deployed to meet the needs of applications in domains such as databases, web service, multimedia, and graphics/visualization. Traditional parallel scheduling techniques such as simple space sharing or coscheduling, which are optimized more for throughput, may not be able to handle the responsiveness needs of these emerging applications. At the same time, leaving it entirely to the native operating system scheduler at each node to make independent decisions can have a detrimental effect on the performance of communicating applications. Considering these different issues, this proposal will explore techniques for efficient management and coordination of the CPU resources across the nodes of a cluster. This research will develop a framework to design cluster scheduling mechanisms taking communication into account together with working sets and I/O information, accomodate the constraints and idiosyncrasies of an application, and adapt dynamically to the system load. The mechanisms will be evaluated with real workloads, using simulation, analytical models and actual implementations. The outcome of this research will take an important step towards making clusters more ubiquitous, multiprogrammed, manageable and user-friendly.