Global Compiler Analysis for Optimizing Tuplespace Communication on Distributed Systems

The generative communication paradigm of parallel programming ooers not only simplicity, but power and ex-ibility. However, the properties of associative access and uncoupled communication that give rise to this power and exibility also lie at the heart of the compiler and run-time system implementation challenges, especially on distributed memory systems. This paper provides concrete steps towards advanced compile-time analysis and optimization of the uncoupled communication of shared tuplespace. Speciically, we present global analysis techniques for detecting common, yet possibly inef-cient, tuple usage patterns at compile time. As part of an optimizing Linda compiler, we have developed and implemented a data ow framework which statically estimates the count of tuples at run-time for each of the tuplespace partitions. We have also designed algorithms for the identiication of shared variable tuples and a class of synchronization tuples. Our empirical ndings show that over 28% of tuplespace partitions in a suite of real application programs never contain more than one tu-ple. These results indicate that there is plenty of opportunity for compile-time optimization of communication of Linda programs, and a global static analysis of Linda parallel programs can indeed provide this information to the optimizer, and thus programmers do not have to rely strictly on run-time and peephole optimizations for achieving good performance.