Highlight

Advancing materials simulations on Summit via RAPIDS

Comparison between the old code and new code, with respect to the time to solution and GPU utilization of the codes on Summit
Four way comparison between the old code and new code, with respect to the time to solution and GPU utilization of the codes on Summit. These numbers were obtained from the profiling of DCA++. We notice a speedup of 15x over the old code, and 47x increase in the GPU utilization of over the old code.

Achievement

  • The authors collaborated with TAU (SciDAC research institute) to profile and analyze DCA++ code.
  • Optimizations to the code was made based on the performance bottlenecks identified by TAU.
  • For this highlight, we observe a 15x speed up over the old code, and upto 47x improvement in the GPU utilization of the code on Summit. 
  • For full scale production run, we observe about 120x speedup over the old code on all 4600 nodes on Summit

Significance and Impact

Collaborating via the RAPIDS Institute, a joint research team from ORNL and the University of Oregon has harnessed TAU’s performance feedback to assist the DCA++ team in exploiting the GPUs on ORNL’s Summit supercomputer. Specifically, TAU has enabled DCA++ developers to improve the code’s performance on the Summit system and increase GPU utilization.

Research Details

  • TAU provides insight in tuning DCA++ execution parameters (e.g. Monte Carlo walkers/accumulators) to run efficiently on Summit and Titan. 
  • Researchers developed ideas to visualize massive amounts of GPU performance data in a scalable way. 
  • TAU facilitates the porting and testing of DCA++ on Summit by integrating DCA into a continuous integration performance system.

Overview

  • DCA++, an ORNL-developed code to simulate correlated quantum materials.
  • Researchers develop optimized algorithm and parallelization strategies for the implementation of new science capabilities in DCA++ using the performance and visualization tool, TAU (Tuning and Analysis Utilities).
  • TAU is a scalable and portable profiling and tracing toolkit for the analysis of parallel programs developed under SciDAC’sRAPIDS Institute.