Replicating data for better performances in X10

Abstract

Linguistic primitives for replica-aware coordination offer suitable solutions to the challenging problems of data distribution and locality in large-scale high-performance computing. The data replication mechanisms that had previously been designed to extend Klaim with replicated tuples are now used to experiment with X10, a parallel programming language primarily targeting clusters of multi-core processors linked in a large-scale system via high-performance networks. Our approach aims at allowing the programmer to specify and coordinate the replication of shared data items by taking into account the desired consistency properties. The programmer can hence exploit such flexible mechanisms to adapt data distribution and locality to the needs of the application, in order to improve performance in terms of concurrency and data access. We investigate issues related to replica consistency and provide a performance analysis, which includes scenarios where replica based specifications and relaxed consistency provide significant performance gains.