PI: Andrew Siegel, MCS

Small modular reactors offer the prospect of affordable electricity production while avoiding some of the traditional limitations that encumber large reactor designs, such as high capital costs and long construction timelines. A principle challenge in the design of these reactors is the absence of extensive experimental validation that exists for current light water reactors in the U.S. nuclear fleet; thus, the design process is heavily dependent on modeling and simulation. However, industry-class computing is based on heavily parameterized, coarse models of reactor phenomena. There is an urgent need for high-resolution calculations that can benchmark these engineering-class simulations.

Objective: To attain efficient execution of Monte Carlo particle transport and CFD algorithms on architectures with relatively deep SIMD units. Thus, our primary hardware interest at present is in the Intel Xeon Phi Knights Landing processor that is already deployed on multiple nodes. To a lesser extent, we are also interested in performance profiling on Xeon processors (thing) and the IBM POWER8 (firestone). Project members from ORNL will be focusing on GPU optimization so there may also be a need for access to resources with GPUs.

Testbed description: To enable high-resolution reactor calculations by integrating Monte Carlo particle transport and computational fluid dynamics—the most accurate numerical methods available for operational reactor modeling—for efficient execution on exascale systems. We are building on a base of applications that have demonstrated high efficiency and excellent scaling on current, petascale leadership-computing levels. This effort will not only provide value to vendors and the broader nuclear community through the generation of highly detailed, benchmark data sets of operational nuclear reactors; it will also analyze important computational motifs under the DOE Exascale Computing Program. The Monte Carlo particle transport method presents significant challenges related to random access of unordered data on hierarchical memory architectures while computational fluid dynamics presents the challenge of achieving high floating-point efficiency on sparse linear algebra problems.