Temperature Aware Load Balancing for Parallel Applications: Preliminary Work

Osman Sarood and Abishek Gupta and Laxmikant V. Kale, "Temperature Aware Load Balancing for Parallel Applications: Preliminary Work",
Parallel Programming Laboratory, Department of Computer Science, University of 
Illinois at Urbana-Champaign, Proceedings of High Performance Power Aware Computing (HPPAC) in IEEE International Parallel and 
Distributed Processing Symposium 2011, May 2011.

@INPROCEEDINGS{Anon1105:Temperature,
AUTHOR="Osman Sarood and Abhishek Gupta and Laxmikant V. Kale",
TITLE="Temperature Aware Load Balancing for Parallel Applications: Preliminary
Work",
BOOKTITLE="The Seventh Workshop on High-Performance, Power-Aware Computing (HPPAC'11)",
ADDRESS="Anchorage, Alaska, USA",
DAYS=16,
MONTH=5,
YEAR=2011,
KEYWORDS="Temperature Aware; Load Balancing; DVFS; Hot Spots; Iterative Applications;
HPC; Cooling Energy Consumption",
ABSTRACT="Increasing number of cores and clock speeds on a smaller chip area implies
more heat dissipation and an ever increasing heat density. This increased
heat, in turn, leads to higher cooling cost and occurrence of hot spots.
Effective use of dynamic voltage and frequency scaling (DVFS) can help us
alleviate this problem. But there is an associated execution time penalty
which can get amplified in parallel applications. In high performance
computing, applications are typically tightly coupled and even a single
overloaded core can adversely affect the execution time of the entire
application. This makes load balancing of utmost value. In this paper, we
outline a temperature aware load balancing scheme, which uses DVFS to keep
core temperatures below a user-defined threshold with minimum timing
penalty. While doing so, it also reduces the possibility of hot spots. We
apply our scheme to three parallel applications with different energy
consumption profiles. Results from our technique show that we save up to
14\% in execution time and 12\% in machine energy consumption as compared
to frequency scaling without using load balancing. We are also able to
bound the average temperature of all the cores and reduce the temperature
deviation amongst the cores by a factor of 3."
}