Low-overhead gang scheduling on workstation clusters

Ku-Jei King, Purdue University

Abstract

Gang Scheduling improves the performance of parallel programs by running all child processes concurrently on their respective nodes. On workstation clusters, the difficulty in implementing gang scheduling lies in the fact that each node runs an antonomous copy of the operating system. In the past, two main approaches have been undertaken by the research community. The first involves handshaking among nodes in order to reach an agreement on overall system state. This exposes synchronization messages to long network delays, and information can become stale very quickly, especially since modern microprocessors have high clock speeds. The second method uses only information that is available locally to estimate the state of the other nodes. Although network delays can be avoided, the heuristic nature of this approach suggests a lack of information accuracy. This thesis presents a novel scheme that offers the precision control of the first method, but like the second method, it requires no explicit handshaking. The key to success is the use of exclusive low-latency communication media with inherent storage capabilities. As results reveal, such a low-overhead scheme provides the flexibility of throttling time-allocation among different processes in a multiprogramming environment. Aggressive settings allow parallel computations to behave as if they are running on a dedicated machine. Less aggressive settings allow fair sharing of system time. In general, the contributions of this thesis reflect how advanced commodity components can be more effectively harnessed to improve parallel performance on workstation clusters.

Degree

Ph.D.

Advisors

Meyer, Purdue University.

Subject Area

Electrical engineering|Computer science

COinS