30th Annual Dayton-Cincinnati Aerospace Science Symposium 
 
 

Performance Optimization of a Structured and Unstructured CFD Code on Commodity
Clusters 

Saurabh Gupta, Dept. of Mechanical Engineering, University of Kentucky
Pavan Kristipati, Dept. of Mechanical Engineering, University of Kentucky
Raymond P. LeBeau, Jr., Dept. of Mechanical Engineering, University of Kentucky
George Huang, Dept. of Mechanical Engineering, University of Kentucky


We present a performance analysis of two in-house codes at the University of Kentucky, the structured two-dimensional code GHOST and the unstructured two/three dimensional code UNCLE, on commodity cluster platforms. The simultaneous optimization of these two codes allows superior methodologies to be identified and tested across different code constructs. Tests were conducted on two different in-house clusters, KFC3 and KFC4, to investigate the effects of performance enhancements on different cluster architectures.

A reasonably detailed discussion of the classification of the problem areas and various restructuring techniques employed to rectify them is presented, with a specific focus on cache performance. Another technique examined is sub-blocking the grid into cache-friendly grid sizes, which for GHOST is done manually on the grid files before uploading while for UNCLE the analogous process is"automated" by using a space filling curve generator. Overall timing of the codes on multiple grid sizes are compared directly to cache activity through use of the Valgrind cache profiler (http://valgrind.kde.org).

Commendable improvements have been achieved in the optimization of GHOST, but the code performance can vary significantly with grid size. The overall performance of the code relative to the original code ranged from improvement of about 25% for smaller grids to a factor of ten improvement for larger grids. Performance improvements for UNCLE are not yet that significant, but based on the GHOST results and the performance characteristics of UNCLE there appears to be considerable room for improvement remaining.