Workshops Archive

Toward OpenSHMEM 2.0: Separating Memory from Execution in PGAS


Workshop: The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM)

Authors: Nathan Wichmann (HPE), Oscar Hernandez (Oak Ridge National Laboratory), Naveen Namashivayam (HPE), Aaron Welch (Oak Ridge National Laboratory), and Stephen Poole (Los Alamos National Laboratory)

Abstract: OpenSHMEM and the wider Partitioned Global Address Space (PGAS) family were built for homogeneous, CPU-centric machines, and they rest on one design choice that modern hardware has outgrown. The agent that issues, orders, and completes communication, which we call the \term{executor}, is bound to the process (the PE) that owns symmetric memory, so that by default one PE is one executor. On a CPU node this costs nothing. On a heterogeneous node, where one process drives many CPU threads, GPU streams, and thousands of GPU blocks, it becomes the central obstacle. We argue that OpenSHMEM 2.0 turns on one decision: \term{split the executor from the process}. Within each executor the model attempts to stay OpenSHMEM 1.x, so a single-executor process is exactly a classic PE and existing programs keep working. This position paper makes the case; it offers no API and no measurements.


Back to The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM) Archive Listing

Back to Full Workshop Archive Listing