July 20, 2017
Frank Mueller, Professor of Computer Science at NC State, has been awarded
$90,000 by Virginia Tech via U.S. Air Force Office of Scientific Research
to support his proposal entitled “A Deep-Learning Approach Towards Auto-Tuning
CFD Codes”. The award runs from June 1,
2017 to February 14, 2018. Co-PIs on the project are Dr. Jack R. Edwards, Jr., Angel Family Professor, and Hong Luo, Professor, in the Mechanical and Aerospace Engineering Department at NC State.
According to the November 2015 TOP500 List,
more than 35% of the computational power of the TOP500 now comes from systems
containing heterogeneous computing devices. However, these computing devices
pose as significant hurdles that impede a scientist’s ability to extract high
performance out of these devices. Mueller’s proposal will study, analyze, and
synthesize deep-learning approaches that can be used for any specific reason
including: performance, power, and energy efficiency.
Abstract - Heterogeneous
computing systems are increasingly becoming the norm in high-performance
computing (HPC). For instance, as of the November 2015 TOP500 List, more than
35% of the computational power on the TOP500 now comes from systems containing
heterogeneous computing devices, e.g., CPUs, GPUs, APUs, Xeon Phis, and even
FPGAs. However, significant hurdles impede a domain scientist’s ability to
extract high performance out of such heterogeneous devices, including (1)
selecting the appropriate algorithm(s) for the target heterogeneous device, (2)
setting the runtime parameters, and (3) configuring the hardware relative to
some evaluation metric, e.g., performance, power, or energy efficiency.
Unfortunately, exploring hardware and software design choices often requires
time-consuming simulations, and while some brute-force auto-tuning support has
been proposed, the results are heuristic that are often narrow in scope, i.e.,
only applicable to a particular problem. This proposal seeks to study, analyze,
and synthesize deep-learning approaches that expose the various parameters as
"knobs" that can be tuned via deep learning to optimize for the
metric of interest, whether it be performance, power, or energy efficiency.