Minimizing Message-Passing Contention in Fragmentation-Free Processor Allocation

Processor allocation, the spatial assignment of a set of processors to each scheduled job, is an important issue in the drive to fully realize the performance potential of massively parallel supercomputers. With the advent of non-contiguous processor allocation strategies , the fragmentation problem has been solved and system utilization has improved. However, message-passing contention is the current performance bottleneck in fragmentation-free processor allocation. In this paper, we present MC, a new processor allocation strategy for mesh-connected parallel computers that (1) is fragmentation-free, (2) yields compact allocations and thus (3) is very successful at minimizing message-passing contention. Moreover, MC is inherently parallelizable, employing just the idle nodes and needing only limited communication. We test MC's performance using a message-passing simulator and workload traces from the San Diego Supercomputing Center. The results show that MC outperforms all other published strategies, reducing average job response time up to 97% and being capable of sustaining higher system loads.