The Effect of Asymmetric Performance on Asynchronous Task Based Runtimes

Abstract

It is generally accepted that future supercomputing workloads will consist of application compositions made up of coupled simulations as well as in-situ analytics. While these components have commonly been deployed using a space-shared configuration to minimize cross-workload interference, it is likely that not all the workload components will require the full processing capacity of the CPU cores they are running on. For instance, an analytics workload often does not need to run continuously and is not generally considered to have the same priority as simulation codes. In a space-shared configuration, this arrangement would lead to wasted resources due to periodically idle CPUs, which are generally unusable by traditional bulk synchronous parallel (BSP) applications. As a result, many have started to reconsider task based runtimes owing to their ability to dynamically utilize available CPU resources. While the dynamic behavior of task-based runtimes had historically been targeted at application induced load imbalances, the same basic situation arises due to the asymmetric performance resulting from time sharing a CPU with other workloads. Many have assumed that task based runtimes would be able to adapt easily to these new environments without significant modifications. In this paper, we present a preliminary set of experiments that measured how well asynchronous task-based runtimes are able to respond to load imbalances caused by the asymmetric performance of time shared CPUs. Our work focuses on a set of experiments using benchmarks running on both Charm++ and HPX-5 in the presence of a competing workload. The results show that while these runtimes are better suited at handling the scenarios than traditional runtimes, they are not yet capable of effectively addressing anything other than a fairly minimal level of CPU contention.