100. phase-based power prediction for heterogenous computing ecosystems

Department: Computer Science & Engineering
Research Institute Affiliation: Graduate Program in Computational Science, Mathematics, and Engineering (CSME)
Faculty Advisor(s): Tajana S. Rosing

Primary Student
Name: Yeseong Kim
Email: yek048@ucsd.edu
Phone: 858-337-9889
Grad Year: 2018

Abstract
The emergence of Internet of Things paradigm is increas- ing complexity and heterogeneity of computing platforms with extremely high demand of energy and performance ef- ficiency. To optimize for a given system power and perfor- mance objective of the computing nodes that significantly differ each other, we have to identify hardware and software behavior over different configurations and even on differ- ent platform architectures. In this paper, we propose P 4 , a new Phase-based Power and Performance Prediction frame- work, which enables to intelligently utilize refined data from performance counters and identify cross-platform applica- tion power consumption at runtime. Unlike existing power estimation techniques, P 4 automatically recognizes distinct application phases to understand machine-independent ap- plication behaviors in a more fine-grained way, without a priori knowledge of the program source code. The frame- work utilizes unsupervised machine learning to identify the phases, which represent key application profiles related to system usage characteristics. Then, neural network-based models perform what-if analysis to predict how performance and power behaviors of application tasks will be changed on a different system. We evaluate the proposed framework on four commercial computing devices consisting of dif- ferent platforms, architectures, and diverse frequency lev- els with 129 industry-standard benchmarks. Our experimen- tal results verify a strong relationship between the extracted phases and the system power consumption, with Coefficient of Variation (CV) of 6.1% for intra-phase power dissimilar- ity. With our phase-based prediction models, the P 4 frame- work can predict the power levels of diverse applications with only 6.8% error for completely different architectures from the ones applications are characterized on.

Industry Application Area(s)
Energy/Clean technology | Internet, Networking, Systems

« Back to Posters or Search Results