Re-examining Scheduling and Communication in Parallel Programs
Abstract: Modern MPPs and NOWs have evolved in ways that affect both the scheduling of parallel applications as well as the communication layer. The presence of a full operating systems upon each processor as well as the need to support interactive users substantially alter the traditional environment. Parallel applications may no longer be executing in the dedicated environment that fast communication layers, such as Active Messages, assume. In this paper we present a simulation-based study of the effects of a non-dedicated environment on parallel applications and investigate one method for reducing the resulting performance impact.
Our results quantify the performance impact of the size of the flow-control window on parallel applications. We investigate increasing the size of this window to ameliorate the effect of various scheduling disturbances. Our results show that additional buffering in the communication layer significantly improves performance in the presence of large scheduling irregularities (e.g., those that occur when parallel applications are locally scheduled) but has a detrimental effect with smaller disturbances (e.g., quantum skew, daemon activity, and interactive users).