Extending a message passing runtime to support partitioned, global logical address spaces

摘要: Partitioned Global Address Space (PGAS) parallel programming models can provide an efficient mechanism for managing shared data stored across multiple nodes in a distributed memory system. However, these are traditionally directly addressed and, applications with loosely-structured or sparse data, determining the location of given element within PGAS incur significant overheads. Applications additional overhead from network latency lookups remote resolution structures. Further, large caching such structures locally incurs space and coherence overheads that limit scaling. We observe matching used by implementations Message Passing Interface (MPI) establish separation between incoming writes where will be stored. In this work, we investigate extending to add layer indirection reads which read, effectively logical addressing.