An implementation of Linda for a NUMA machine

After briefly reviewing the implementation strategies for Linda on shared-memory, distributed-memory, and LAN architectures, we discuss a Linda implementation for the CRAY T3D and CRAY T3E. We identify some significant ways in which the architecture of these machines differs from those listed - in essence the CRAY T3D and CRAY T3E can be viewed as instances of an architecture somewhere between shared-memory and distributed-memory. As a result, a new implementation strategy was called for. We describe two approaches that we explored and lay out the more effective one in some detail. We conclude with performance data for point-to-point coordination operations that show that our Linda implementation generally outperforms MPI and PVM on the CRAY T3D and is comparable to MPI on the CRAY T3E. This work thus extends a series of results that have demonstrated `high-level' coordination languages, like Linda can be implemented on modern multiprocessors without significantly compromising performance relative to low-level coordination via message passing.