Optimizing Software Shared Memory Communication with InfiniBand

Implementations of algorithms based on shared memory are widely used in the area of High Performance Computing. The concept of cluster computing, however, does not imply the availability of a globally shared memory across nodes in that cluster. Therefore, several concepts for an emulation of a globally shared memory, called Software Distributed Shared Memory (SDSM), exist. However, these concepts did not achieve widespread acceptance because a significant part of the overall execution time is spent on the emulation of the shared memorv region. With the introduction of highspeed interconnect-technologies, e.g. Myrine¹ or InfiniBand², higher bandwith and lower latencies have become available. However, special inter-process communication methods like direct access to other nodes' memory are especially useful for the further development of existing SDSM concepts. Using this techniques, existing implementations can be greatly simplified and optimized. This paper discusses some common problems of existing SDSM-protocols and then presents solutions for these issues by taking advantage of InfiniBand specific communication patterns.