Privatization and Distribution of Arrays A Preliminary Proposal

In today's high performance NUMA (Non-Uniform Memory Architecture) multiprocessors with memory hierarchy or distributed memory, the partition and distribution of data associated with parallel computations a ect the amount of parallelism that can be exploited and the amount of data movement in the system. The objective of this research is to study and evaluate compile time data management techniques to enhance parallelism and to improve locality of memory reference for large scienti c programs written in Fortran. Our rst step is to reduce the amount of shared data through privatization. Privatization is a technique that allocates a separate copy of a shared variable in the private storage of each processor such that each processor can access a distinct instance of the variable. Privatization can enhance inherent parallelism of a program by eliminating memory-related antiand output dependences. It can also improve the locality of references since accessing a private variable is inherently local and communication free. We present our algorithm for array privatization and the result of our experiment on the e ectiveness of the algorithm. For the remaining shared data, we introduce a new concept: placement matrix, and show its application in deriving data alignment and data decomposition to reduce communication. We also incorporate the ratio of communication to computation in our evaluation of di erent data partitions. The work is continuing on heuristics for data distribution and the implementation of the tools.