PARADYN

While employed at SwRI in San Antonio, TX, I developed a parallel, dynamic optimization engine driven by a genetic algorithm for the purposes of optimizing the performance of computationally intensive simulations/analyses at runtime. Compilers are excellent at performing static analysis to deliver the most optimize code possible based on user hints and/or profiling results. However, what if the computational characteristics of a simulation/analysis change dramatical as the simulation runs? Then, any sort of compile-time optimization at the beginning may well be for naught. Enter PARADYN, a distributed optimizer targeted at heterogeneous clusters. PARADYN would:

• Maintain a population of the best formulations of a code base from a computational optimization standpoint, and run them all simultaneously on the cluster.

• Monitor the per-timestep execution time of all candidate solutions, speculatively executing the ones that maintained high performance, and culling the poorly performing ones and replacing the others with new solutions obtained from the genetic algorithm