A Deep Learning Approach towards Auto Tuning CFD Codes
Abstract
Heterogeneous computing systems are increasingly becoming the norm in high-performance computing (HPC). For the June 2018 TOP500 List, the majority of computational power on the TOP500 comes from systems containing heterogeneous computing devices, e.g., CPUs, GPUs, APUs, and Xeon Phis. However, significant hurdles impede a domain scientists ability to extract high performance out of such heterogeneous devices, including (1) selecting appropriate algorithm(s) for the target heterogeneous device, (2) setting runtime parameters, and (3) configuring hardware relative to some evaluation metric, e.g., performance, power, or energy efficiency. Furthermore, given the diversity of HPC systems, domain scientists want their software codes to be portable across many computing systems and to understandably have some measure of future-proofing of their software codes, even as the underlying hardware continues to rapidly evolve. Thus, our project studies, analyzes, and synthesizes machine-learning and deep-learning approaches that expose the parameters as knobs and tune them dynamically during the simulation so as to optimize for the metric of interest, whether it be performance, power, energy efficiency, numerical accuracy, or fidelity of the flow physics.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 18, 2018
- Accession Number
- AD1060773
Entities
People
- Wu-chun Feng
Organizations
- Virginia Tech