PERFORMANCE IMPLICATIONS OF COMMUNICATION MECHANISMS IN ALL-SOFTWARE GLOBAL ADDRESS SPACE SYSTEMS

Global addressing of shared data simplifies parallel programming and c omplements message passing models commonly found in distributed memory machines. A number of programming systems have been designed that syn thesize global addressing purely in software on such machines. These s ystems provide a number of communication mechanisms to mitigate the ef fect of high communication latencies and overheads. This study compare s the mechanisms in two representative all-software systems: CRL and S plit-C. CRL uses region-based caching while Split-C uses split-phase a nd push-based data transfers for optimizing communication performance. Both systems take advantage of bulk data transfers. By implementing a set of parallel applications in both CRL and Split-C, and running the m on the IBM SP2, Meiko CS-2 and two simulated architectures, we find that split-phase and push-based bulk data transfers are essential for good performance. Region-based caching benefits applications with irre gular structure and with sufficient temporal locality, especially unde r high communication latencies. However, caching also hurts performanc e when there is insufficient data reuse or when the size of caching gr anularity is mismatched with the communication granularity. We find th e programming complexity of the communication mechanisms in both langu ages to be comparable. Based on our results, we recommend that an idea l system intended to support diverse applications on parallel platform s should incorporate the communication mechanisms in CRL and Split-C.