SC15 Austin, TX

Designing High Performance and Energy-Efficient MPI Collectives for Next Generation Clusters

Student: Akshay Venkatesh (Ohio State University)
Advisor: Dhabaleswar Panda (Ohio State University)
Abstract: The past decade has witnessed a notable change in HPC landscape with the proliferation of accelerators/co-processors.The presence of processing cores and memory of a different nature has had an adverse impact on the MPI and PGAS programming models that are predominantly used for developing scientific applications.As accelerators/co-processors are primarily available as external PCIe devices, the external cores and memory units render systems heterogeneous.From the perspective of collectives communication routines, where hundreds to thousands of cores are likely to be involved, traditional assumptions such as uniform processing and communication costs within the system are rendered invalid.As the exascale report identifies communication and energy optimization among the foremost challenges in reaching the exaflop mark, this work addresses: 1.the performance of collective algorithms by distinguishing heterogeneous communication path costs and 2.the energy of both collective and point-to-point operations by generating rules that apply energy saving strategies at the fundamental underlying communication algorithms.

Summary: pdf
Presentation: pdf
Poster: pdf

Doctoral Showcase Index