Optimizing the Energy Usage and Cognitive Value of Extreme-Scale Data-Analysis Approaches
James Paul Ahrens, David Rogers, Roxana Bujack, Anna Berres, Wu-chun Feng, Vignesh Adhinarayanan, Colin Ware, Francesca Samsel, Gregory D. Abram, Terece L. Turton · 2017
Scientific discovery at the extreme scale is a unique technical challenge, requiring the reduction of massive amounts of data into compact analysis products that capture key scientific insights. This analysis process needs to occur under extreme-scale computational constraints including minimizing 1) data movement, 2) energy usage and 3) storage usage. Put simply, extreme-scale computing platforms are to achieve a three orders-of-magnitude increase in computational performance while consuming only two times the electrical power of current platforms. Data movement costs will dominate energy usage at this scale, so the HPC community expects extreme scale analysis algorithms will be utilized to reduce simulation results in-situ – that is, during the simulation run. This reduction will occur, broadly speaking, via some type of adaptive sampling, such as signal, statistical or feature-based sampling. For this project, we considered the computational platform and the scientists using the platform as a system to be optimized. Our goal was to maximize scientific insight from sampled results while minimizing power. We proposed to investigate how changes in our sampling algorithms – necessary because of exascale power constraints – impacted the cognitive value of the resulting data. Our goal was to provide tools and methods for scientists and others to explore, understand, and utilize the high dimensional tradeoff space embodied in the inputs or controls to these workflows, e.g. data sampling approach, visualization pipeline, job architecture, and visualization approach. Our work progressed from basic initial experiments through more detailed explorations, and resulted in both research results and a set of tools that can be used to explore these areas further. A key component of the success of this project was our approach of integrating the expertise of a diverse set of researchers. Experts from computing, perception, cognition, energy and power, and color theory came together to devise and execute experiments and develop tools to address the complex problem of maximizing scientific insight under constraints. The ECX team developed a close, effective working relationship across four disparately-located institutions, developed open source software together, collaborated on papers, and crowdsourced user testing methods. This team and its work is the basis for the successfully funded proposal A Continuously-Running, Asynchronous, Sampling Engine for the Perceptual and Cognitively-driven Visual Analysis of Massive Scientific Data.