On the Performance of Parallel Processing in Dynamic Resource Sharing Systems

Parallel processing has the potential of significantly decreasing the service time for a single computational task. Meanwhile, as each task occupies more resources, the number of simultaneously supported tasks declines. This tradeoff is interesting when resources are accessed by many users in a dynamic way, like in the case of cloud or fog computing. In this paper, we evaluate how the level of parallelization and the eventual overheads affect the response time in these dynamic resource sharing systems. We show the counterintuitive finding that even when parallelization has no overhead, the allocation of all resources to a task is suboptimal if the service times have large coefficient of variation. Moreover, we evaluate the scalability properties, and provide guidelines for the optimal level of parallelization under different types of overhead.