Resource management is a key area in the drive to fully realize the performance potential of parallel and distributed systems. For the task of assigning a set of processors of a massively parallel processing (MPP) system to a given job, various processor allocation strategies have been proposed in the research community and are in use at supercomputing sites. With the advent of the class of non-contiguous allocation strategies, the allocation performance bottleneck shifted from fragmentation to message-passing contention.

This paper presents a method to estimate and minimize contention incurred by non-contiguous allocation strategies. Our approach is to analyze the spatial layout of dispersed nodes. Our contribution is a set of dispersal metrics that predict contention well and that are efficient to implement for a variety of interconnection topologies. We put the dispersal metrics to the test by comparing their contention estimates with measurements taken from a message-passing simulator. Our analysis and experiments consider different topologies of machines, a wide range of communication patterns and different workloads. Because our results show very high correlations between dispersal metrics and contention, we conclude that dispersal metrics have the potential to help evaluate and improve processor allocation strategies.