高生産な並列プログラミング環境を実現するための並列ランタイムシステムの設計

Abstract

To realize highly productive parallel programming environments, newly designed parallel runtime system based on software distributed shared memory is proposed. It can provide a seamless parallel programming model from intra-node multithread programming to inter-node parallel programming. It overcomes the data inconsistent problem caused by multiple threads in a user program at remote page fetching. It provides node-shared data address space with a traditional relaxed memory consistency model as well as high-speed ad-hoc data copy interface between node-shared data and node-local data. User can easily extend their existing C, OpenMP and OpenACC programs to node parallel programs incrementally by directive-based programming API on this runtime system.