On the benefits and limitations of dynamic partitioning in parallel computer systems

Abstract

In this paper we analyze the benefits and limitations of dynamic partitioning across a wide range of parallel system environments. We formulate a general model of dynamic partitioning that can be fitted to measurement data to obtain a sufficiently accurate quantitative analysis of real parallel systems executing real scientific and/or commercial workloads. An exact solution of the model is obtained by employing matrix-geometric techniques. We then use this framework to explore the parallel system design space over which dynamic partitioning outperforms other space-sharing policies for a diverse set of application work-loads, quantifying the significant performance improvements within these regions. Our results show that these regions and the performance benefits of dynamic partitioning are heavily dependent upon its associated costs, the system load, and the workload characteristics. We also identify the regions of the design space over which dynamic partitioning performs poorly, quantifying the performance degradation and illustrating forms of unstable thrashing.KeywordsSystem LoadLinear SpeedupParallel Computer SystemDynamic PartitioningWorkload CharacteristicThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.