Energy-to-Solution : a today's metric for tomorrow's concerns

Abstract : « Exascale » is the last buzz word in the high performance computing world. Big chip founders and supercomputers vendors are – evidently – the first Moore’s law believers, and they will probably achieve 1018 [GF/s] on the synthetic benchmark HPL within a decade. The consequences for real applications are still to be measured. Today’s largest machines consume a dozen of megawatts of electricity while the computational efficiency of a real application is less than 5%. Following the opposite trend, energy becomes the issue in the development of the aging CMOS technology on the road to exascale. I'll start with a short historical survey of the energy and computational efficiencies during the last 15 years of chip development and the consequences of the so-called « Memory Wall ». Then, I'll introduce a new metric, the Energy-To-Solution quantity by comparing the needs of the applications with today’s clusters and supercomputers capabilities. A real life example is shown. Three real applications have been run on three 2012 processors. Power consumption was measured during execution. The astonishing results lead to a new decision scheme based on the energy efficiency to choose an existing (or a new) machine to run a scientific application in the most efficient way. This decision scheme should also act as a lighthouse for future chip development.

Biography : Vincent Keller is a computer scientist born in 1975 in Lausanne, Switzerland. He received his Master degree in Computer Science from the University of Geneva (Switzerland), then his PhD degree in 2008 from the École Polytechnique Fédérale de Lausanne (EPFL), Switzerland in the HPCN and HPC Grids fields. He is application analyst in the Center for Advanced Modeling Science (CADMOS) at École Polytehnique Lausanne. Dr. Vincent Keller is the author or coauthor of more than 30 peer-reviewed papers and one book (HPC@GreenIT, Springer Verlag, Heidelberg, Germany, October 2009). His research interests are in HPC applications analysis, Grid and cluster computing and energy efficiency of large computing ecosystems.