The International Conference for High Performance Computing, Networking, Storage and Analysis
GMT: Enabling Easy Development and Efficient Execution of Irregular Applications on Commodity Clusters.
Authors: Alessandro Morari (Pacific Northwest National Laboratory), Oreste Villa (NVIDIA Corporation), Antonino Tumeo (Pacific Northwest National Laboratory), Daniel Chavarria (Pacific Northwest National Laboratory), Mateo Valero (Universitat Politecnica de Catalunya)
Abstract: In this poster we introduce GMT (Global Memory and Threading library), a custom runtime library that enables efficient execution of irregular applications on commodity clusters. GMT only requires a cluster with x86 nodes supporting MPI. GMT integrates the Partititioned Global Address Space (PGAS) locality-aware global data model with a fork/join control model common in single node multithreaded environments. GMT supports lightweight software multithreading to tolerate latencies for accessing data on remote nodes, and is built around data aggregation to maximize network bandwidth utilization.