Since the message delivery time of asynchronous applications is unpredictable, each asynchronous process must probe (or poll) the network for new messages regularly. If polling is carried out too aggressively, a majority of the probes will be unsuccessful. However, the cost for probing the network is similar to the cost of communication operations. Thus, optimizing the message reception behavior has a significant impact on the performance of the application. This paper studies this problem, and develops a cost model to optimize the polling frequency. We also develop strategies to optimize the message delivery behavior, and study their impact on the performance of a parallel discrete event simulator.
