Performance and correctness of software for large-scale heterogeneous systems

Richard Vuduc
Georgia Institute of Technology
-------------------------------------------------------

Abstract: 

We give an overview of the major performance and correctness
issues we faced while designing and implementing the first
highly-scalable simulation of deformable red blood cells in plasma.
The resulting code has scaled to 200k cores and achieved 0.7
Petaflop/s of sustained performance, and furthermore shown to work
well on GPU-accelerated clusters; for these demonstrations, the
software received the 2010 Gordon Bell Prize. However, the road there
was paved with scalability missteps, numerical accuracy issues, and
other software engineering challenges typically encountered in
developing HPC codes on bleeding-edge machines. This talk summarizes
these challenges and, with a little luck and suitable interpretation,
suggests ways in which the EC^2 community might help to improve HPC
software development productivity.