News Archives

RSS Feed

  • UNM
  • >Home
  • >News
  • >2014
  • >December
  • >[Colloquium] Beyond FLOPS: realistic performance evaluation of extreme scale applications

[Colloquium] Beyond FLOPS: realistic performance evaluation of extreme scale applications

December 9, 2014

Watch Colloquium: 

MOV FILE

  • Date: Thursday, December 4, 2014
  • Time: 11:00 - 12:15 PM
  • Place: Dane Smith Hall 125

Speaker: Patrick M. Widener, Ph.D. Sandia National Laboratories

Abstract: The intersection of three trends is complicating the design and measurement of new high-performance computing (HPC) systems. Increasing size and component counts of modern and projected systems are presenting unpalatable reliability implications for applications; coupled analysis workflows are integrating traditional HPC resources into solutions for new problem domains; and the size and complexity of production computational science applications makes their evaluation in the presence of new software and hardware capabilities difficult. In this talk I will describe how research undertaken by the 9-Lives SNL/UNM collaboration is reconsidering benchmarking and evaluation of large-scale computations for the coming era of extreme-scale computing where fault-tolerance is a first-class design objective, in-situ analysis will be commonplace, and understanding the tradeoffs posed by new system and runtime stacks will be crucial. Along the way, I will present some preliminary results upon which we hope to expand as part of upcoming funding proposals.

Bio:Patrick M. Widener is a Principal Member of Technical Staff at Sandia National Laboratories. He is co-/author of numerous peer-reviewed conference and journal articles, has held research faculty appointments in Emory University's Department of Biomedical Informatics and the University of New Mexico's Department of Computer Science, and has several years experience as a professional software developer. His research interests include operating systems, fault-tolerance, system software for I/O and storage, and data/metadata management for large-scale coupled applications and services.