An Incremental Parallelization Approach Applied to the ORNL/NRC FAVOR Code

Abstract

Parallelizing a domain-specific production code with thousands of lines that is developed over several years is a daunting task. Also, throwing away the existing serial code completely to design a parallel algorithm from scratch is not always a viable solution. Hence, the parallelization task of the production code is very conservatively approached. Ideally, one wishes to morph the serial code to make it compute in parallel, so that much of knowledge built over years in the form of serial code is retained and the performance gain due to parallel computing is also achieved. Instead of starting from a clean slate, we start at the end product, i.e. the serial code and begin by questioning that if the serial code were to be considered as parallel code and when executed as a parallel program, what problems arise? Addressing each of the arising issues individually, we transform the serial code into the required parallel code. Since the issues are incrementally addressed, we refer to this strategy as Incremental Parallelization Approach (IPA). We demonstrate the applicability of IPA, by parallelizing over 25000 lines of Probabilistic Fracture Mechanics (PFM) module code of Fracture Analysis of Vessels Oak Ridge (FAVOR) code that was developed for the Nuclear Regulatory Commission (NRC) by the HSST program at Oak Ridge. By doing so we reduced the runtime of FAVOR PFM module from several hours to minutes, without any loss of accuracy in the computed result. In this paper we discuss our experience in parallelizing the FAVOR’s PFM module using the IPA.