Efficient Computing

Abstract

As computer systems become parallel and heterogeneous, applications still rely on entrenched control-flow based, CPU-centric programming and execution model. The result is redundant computation, unnecessary data movement and under-utilized hardware. To maximize the performance of parallel, heterogeneous computer architectures, we need to eliminate redundant computation, avoid unnecessary copies and perform computation in the most efficient resource. In this talk, Hung-Wei will share his experiences to achieve efficient computing through revisiting the design of computer architecture, programming and execution models, compilers, system software and storage systems. Hung-Wei will demonstrate using the data-triggered threads model can eliminate redundant computation and achieve 1.46x speedup for SPEC2000 applications. Hung-Wei will also present the Gullfoss system that provides efficient data transfer inside heterogeneous computer systems to accelerate GPU applications by 46%. Hung-Wei will also show that by creating application objects inside storage device, we can speedup applications by 1.32x.