Title: Light Weight Processing for MPI

Speaker: Arun Rodrigues, Notre Dame University

Location: Building 980, Room 90 (NM), Building MO52, Room 165 (CA)

Brief Abstract: Modern high performance network interfaces use programmable NIC processors to offload much of their MPI processing. Unfortunately, a large portion of the MPI processing involves searching lists of MPI request records. These lists can grow quite long and because of the simplicity of the NIC processors can become a bottleneck.

This talk explores a possible solution - replacing the conventional NIC processor with a Light Weight Processor (LWP). Light Weight Processing combines the related techniques of Processing-in-Memory, wide word ALUs, and hardware multi-threading. An LWP has naturally low latency access to memory, so it traverses linked lists well. Wide memory accesses and a vector ALU allow multiple list entries to be searched rapidly. The LWP's hardware support for multi-threading and fine-grained synchronization enable simultaneous queue traversals - harnessing the increased memory bandwidth. Finally, LWPs are less complex, so their area and cost will be minimal. In short, LWPs are a natural fit to the needs of an MPI-accelerating NIC processor.

CSRI POC: Keith Underwood, (505) 284-9283