Low latency remote memory access on the Intel Paragon

We describe a cooperative project between Pacific Northwest Laboratory and Intel on providing low latency communications on the Intel Paragon. Our interest is in developing modules that are tailored to specific types of communication and are optimized for these cases. The modules described are implemented as simple extensions to the Paragon operating system. The first extension supports remote reads, writes, and accumulates on subsections of matrices. This module is intended to be used where the communication pattern is not known until run time. The second extension is designed to support low latency communication for short messages in a pipelined fashion. Finally, an interface for doing highly structured communication where the patterns of communication are well understood ahead of time is described. This is intended to be used in global communications such as global reduction.