Program-based static allocation policies for highly parallel computers

Static space sharing is a promising processor allocation strategy for highly-parallel computers. Parallel jobs or tasks are allocated fixed distinct subsets of processors. Thread scheduling can be carried out directly by the applications so as to reduce its cost, interference among jobs is reduced, and a wide range of compiler and runtime optimizations, including static data distribution and binding, are possible. The fundamental problem with static space sharing is processor fragmentation. In this paper we study and compare several no-folding and unlimited-folding program-based static space sharing processor allocation policies. A job is always allocated the number of processors it requests in the no-folding schemes, whereas it can be allocated any number of processors that does not exceed its request in the unlimited-folding policies. The results of a detailed simulation study show that unlimited folding is superior to no folding under medium to high system loads when the speedup curves of applications are significantly sublinear. However, no folding is superior under most or all loads, depending on workload characteristics, when the applications are highly efficient. The unlimited-folding policies suffer more processor fragmentation under most loads, but they exploit the increase in efficiency that typically results when a job executes on fewer processors. In addition, we study schemes based on the assumption that the execution times of jobs when they execute on number of processors they request are known. The results show that giving priority to shorter jobs can result in significant improvement in mean response times