Workshop: The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM)
Authors: Nicholas Cassarino, Swann Perarnau, Jonathan Ozik, Nicholson Collier, Jamie Cook, and Kevin Brown (Argonne National Laboratory)
Abstract: Parallel discrete-event simulation (PDES) and agent-based modeling and simulation (ABMS) are two important modeling and simulation paradigms used in domains such as network simulation, epidemiology, and transportation. Both must coordinate distributed state as simulation time advances, yet common approaches in each paradigm sacrifice either performance through blocking synchronization or scalability and simplicity through rollback mechanisms to do so. Moreover research in each field has largely progressed in isolation.
We present a partitioned distributed versioned array that can serve as a common global distributed memory layer in both PDES and ABMS for more flexible synchronization schemes. By allowing each rank to maintain multiple versions of a shared datastructure available at the same time, our approach enables unbalanced workloads to drift out of lockstep without rollbacks or neighborhood barriers. We demonstrated increased performance using this scheme on common benchmarks from both PDES and ABMS paradigms, against state of the art runtimes.
Back to The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM) Archive Listing
Back to Full Workshop Archive Listing