OSPRI: An Optimized One-Sided Communication Runtime for Leadership-Class Machines

Abstract—Partitioned Global Address Space (PGAS) program-ming models provide a convenient approach to implementing complex scientific applications by providing access to a large, globally accessible address space. This paper describes the design, implementation and performance of a new one-sided communica-tion library that attempts to meet the needs of PGAS models, par-ticularly Global Arrays, but hopefully also PGAS languages like UPC and CAF. In this work, we describe a new communication runtime for PGAS models such as GA, termed OSPRI (One-Sided PRImitives). OSPRI presents several changes in architecture from conventional one-sided communication systems that make it better suited for emerging leadersip class machines. We describe the implementation of the the IBM Blue Gene/P target for OSPRI and demonstrate significant improvements in latency, bandwidth, and scalability over tuned ARMCI and GA implementations on this system. The performance and scalablity of this library validate the design choices and should provide useful insight for implementers of related communication middleware