Speedup of MCNP(X) parallel KCODE execution via communication algorithm development and Beowulf Cluster optimization

Abstract

This thesis examines two methods for speeding up MCNP KCODE calculations. The first approach is assembly of a low cost Beowulf Cluster for parallel computation. The first half describes the MIT Nuclear Engineering Department's Beowulf Cluster, Echelon, as a prototype. Echelon is a low cost, 30 CPIJ, disk-less Debian Linux Beowulf Cluster. A user's manual is documented and a series of tips given on how to build a similar cluster, and a method described for optimally using the Monte Carlo N Particle (MCNP) Code on this cluster. The focus is on the complex software setup required but also touches upon hardware optimization and cost. The second method is optimization of the parallel KCODE communication algorithm in MCNP. A new patch for the parallel MCNPX KCODE communication algorithm is examined. The new patch facilitates the local execution of KCODE cycles, greatly eliminating interprocess communication. Combined with numerous minor changes to MCNPX the new patch file can speed up MCNPX KCODE calculations by an order of magnitude on a Beowulf Cluster using 60 CPUs. The conceptual design of the new algorithm and the actual changes made to the code are described including showing how an MCNPX user can take advantage of the new features in practice, and providing timing results comparing the new and old algorithms. Combined, the Echelon system and the new communication algorithm speeds up MCNP KCODE execution by 20 fold in parallel and nearly 30 fold with multiple jobs. This allows advanced criticality and complex burnup calculations that previously required a day to be done in an hour. This significant performance gain opens a whole new level of problem complexity for a relatively low cost.