Abstract: "Scheduling in many multiprocessor operating systems is based upon the use of a single priority queue of runnable tasks. This shared queue can become a system bottleneck due to contention among the processors. In this study we address contention for the shared system queue and develop...

Full description

Bibliographic Details

Main Authors:	Nelson, Randolph D. (Author), Squillante, Mark S. (Author)
Format:	Book
Language:	English
Published:	Seattle, Wash. 1989
Series:	University of Washington <Seattle, Wash.> / Department of Computer Science: Technical report 89,11,6
Subjects:	Multiprocessors Operating systems (Computers) Production scheduling
Summary:	Abstract: "Scheduling in many multiprocessor operating systems is based upon the use of a single priority queue of runnable tasks. This shared queue can become a system bottleneck due to contention among the processors. In this study we address contention for the shared system queue and develop scheduling policies to reduce anticipated response time and load balancing problems. These problems are analyzed within the context of a unifying queueing theoretic scheduling model, which is solved using matrix-geometric techniques. Our analysis illustrates that contention for the shared system queue is significant and increases with the number of processors in the system We show that one of our policies, called autonomous scheduling, reduces contention for the system queue and thus provides better system performance than typical single queue scheduling. We also show that another of our policies, called cooperative scheduling, effectively eliminates contention for the shared system queue and consistently outperforms autonomous scheduling. In particular, the cooperative policy provides a greater maximal throughput, a lower mean response time and a smaller response time variance relative to the autonomous policy, and these differences widen with increasing system load, scheduling overhead and number of processors.
Physical Description:	20 S.