Towards unification of accelerated computing and interconnection for extreme-scale computing

Abstract

Heterogeneous clusters using accelerators are widely used for high-performance computing system. In such systems, the inter-node communication among accelerators becomes bottleneck due to the data transfer between the accelerator and the host. To eliminate this overhead, we have been developing a novel communication system realizing direct communication among accelerators over computation nodes under the HA-PACS (Highly Accelerated Parallel Advanced system for Computational Sciences) project. Also we are investigating high-level parallel programming language, and several practical application programs on our concept, as well as studying the enhancement of TCA and developing system software stack in the CREST project.