Maximizing Performance Metrics for Science and Engineering Applications

1. Application Scaling

2. Application Scaling • Doug Kothe, ORNL • Paul Muzio, CUNY • Jonathan Carter, LLBL • Bronis de Supinski, LLNL • Mike Heroux, Sandia • Phil Jones, LANL • Brent Leback, The Portland Group • Piyush Mehrotra, NASA Ames • John Michalakes, NCAR • Nir Paikowsky, ScaleMP • Galen Shipman, ORNL • Trey White, ORNL

3. Application Scaling • Are there more appropriate domain-specific performance metrics for science and engineering HPC applications available then the canonical percent of peak? • Or parallel efficiency and scalability? • If so, what are they? • Are these metrics driving for weak or strong scaling or both?

4. Application Scaling • What is the role of local (node-based) floating point accelerators (e.g., cell, GPUs, etc.) for key science and engineering applications in the next 3-5 years? • Is there unexploited or unrealized concurrency in the applications you are familiar with? • If so, what, and where is it?

5. Application Scaling • Should applications continue with current programming models (Fortran, C, C++, PGAS, etc.) and paradigms (e.g., flat MPI, hybrid MPI/OpenMP, etc.) over the next decade? • If not, what needs to change? • How might HPC system-attribute priorities change over the next decade for the science and engineering applications you are familiar with?

6. Application Scaling • Attributes to consider are: • node peak flops • mean time to interrupt • wide-area network bandwidth • node memory capacity • local storage capacity • archival storage capacity • memory latency • interconnect latency • disk latency • interconnect bandwidth • memory bandwidth • disk bandwidth