Comparison of Processor Allocation Policies for Parallel Systems

The increasing use of parallel systems has led to the development of a number of multiprogrammed processor allocation policies. This paper analyzes the following four policies that have previously been shown to have high performance under speciic workloads: adaptive static partitioning (ASP), dynamic rst-come-rst-serve (FCFS), preemptive smallest available parallelism rst (PSAPF), and spatial equipartitioning (EQS). The results in this paper are derived for a general workload model that includes general distribution of available job parallelism, controlled correlation between cumulative processing demand and available parallelism, general demand distribution per class of jobs in the correlation model, and general deterministic job execution rates that represent synchronization and communication overheads as well as load imbalance for parallel programs. Under the assumption that jobs can dynamically and eeciently redistribute their work across the processors allocated to them previous interpolation approximations are used to estimate the mean response times of EQS and FCFS, and new interpolations are derived and validated for the mean response times of ASP and PSAPF. The interpolation approximations provide approximate mean response time formulas for each policy that directly yield key determinants of relative policy performance. The key determinants are used to delineate regions of the workload parameter space over which each of ASP, FCFS, EQS, and PSAPF performs best. The delineation provides a uniication and generalization of previous results.