A Data Driven Model for Tolerating Communication Delays

Jacob Sorensen, Scott B. Baden

Abstract

Hiding communication latency is a difficult challenge due to the high software overheads incurred. We discuss Thyme, a data-driven programming model and run time library that manages communication pipelining and scheduling through task graph, actor-like execution. Thyme frees the user from many low-level system dependent details involved in masking communication, easing the reformulation of applications to tolerate delays. We demonstrate results with some practical applications and discuss compiler support using the ROSE source-to-source translator.