It’s one of the most talked about issues in the HPC community: the sheer amount of power required by the coming generation of supercomputers. The figures are staggering. Built with today’s technology, the yearly power bill for an exascale machine would be in the neighborhood of a half-billion dollars.
Hitting the exascale target will require a cohesive yet multi-faceted effort. There are programs in place working on the various parts: the memory wall, the power wall, the programming challenge and so forth. The EU-funded ADEPT project is exploring the energy-efficient use of parallel technologies.
ADEPT, which stands for ADdressing Energy in Parallel Technologies, has two main parts: 1) help HPC software developers exploit parallelism for performance, and 2) assist embedded systems engineers with managing energy usage.
The end goal is the creation of a tool that will help users model and predict the power consumption and performance of their code.
In essence, ADEPT combines the talents of the HPC and embedded communities, drawing on the unique strengths of each sector: parallelization on the HPC side and energy management on the embedded side. It’s an opportunity for each discipline to learn from the other.
“The strength of the HPC world lies primarily in software application parallelisation: concurrent computation is used to speed up the overall time an application requires run to completion,” remarks the project website. “As a result, HPC software developers are also experts in parallel performance analysis and performance optimisation.”
It continues: “The embedded systems sector excels in managing energy usage because it is often constrained by fixed power and energy budgets.
“The strengths of one sector are the relative weakness of the other: power management and power efficiency in HPC are in their infancy, but they are becoming increasingly important with HPC systems requiring more and more power; and while continuing to be constrained by energy and power budgets, recent advances in low-power multi-core processors have widened the choice of hardware architectures for embedded systems and are increasingly forcing embedded programmers to use parallel computing techniques that are more familiar to HPC programmers.”
The program reflects the new energy-aware HPC paradigm. Between the high-cost of energy and its environmental implications, high-performance computing, and computing in general, have by necessity begun to emphasize performance-per-watt over pure performance. To this end, ADEPT aims to provide a better understanding of how parallel software and hardware use power. Reducing the power usage of large-scale applications will help make exascale systems more feasible from an economic and environmental standpoint.
ADEPT is scheduled to run three years, commencing on Sept. 1, 2013. The program is facilitated by Edinburgh Parallel Computing Centre, the supercomputing center at the University of Edinburgh, and includes partners Uppsala University (Sweden), Alpha Data (UK), Ericsson AB (Sweden) and Ghent University (Belgium).