Space sharing job scheduling policies for parallel computers

The distinguishing characteristic of space sharing parallel job scheduling policies is that applications are allocated non-overlapping processor subsets. The interference among jobs is reduced, the synchronization delays and message latencies can be predictable, and distinct processors may be allocated to cooperating processes so as to avoid the overhead of context switches associated with traditional time-multiplexing;The processor allocation strategy, the job selection criteria, and workload characteristics are fundamental factors that influence system performance under space sharing. Allocation can be static or dynamic. The processor subset allocated to an application is fixed under static space sharing, whereas it can change during execution under dynamic space sharing. Static allocation can produce more predictable run times, permits a wide range of compiler optimizations (e.g., static data distribution and binding), and avoids the processor releases and reallocations associated with dynamic allocation. Its major problem is that it can induce high processor fragmentation;In this dissertation, alternative static and dynamic space sharing policies that differ in the allocation discipline and the job selection criteria are studied. The results show that significantly superior performance can be achieved under static space sharing if applications can be folded (i.e., allocated fewer processors than they requested). Folding typically increases program efficiency and can reduce processor fragmentation. Policies that increase folding with the system load are proposed and compared to schemes that use unconstrained folding, no folding, and fixed maximum folding factors. The adaptive policies produced higher and more stable system utilization, significantly shorter mean response times, and good fairness curves. However, unconstrained folding resulted in considerably more severe processor fragmentation than no folding. Its advantage is that it exploits the efficiency improvement that typically results when an application is allocated fewer processors. Consequently, it can produce shorter mean response times than no folding under medium to heavy loads;Also because of this efficiency improvement, dynamic policies that reduce waiting times by executing a large number of jobs simultaneously are more promising than schemes that limit the number of active jobs. However, limiting the number of active applications can be the superior approach when folding does not improve application efficiency