Resource Intelligent VPU Utilization Leading to an Effectively Threaded System (RIVULETS)

To improve performance of Army CFD codes on new HPC architectures using MIC co-processors, we propose the RIVULETS tool which uses hybrid algorithms and automatic tuning to decrease CFD run-time and improve numerical performance. Existing CFD codes are parallelized using MPI, a highly effective approach for distributed architectures with large quantities of system memory and relatively simple serial processors. However, new HPC systems based on MIC co-processors have fundamentally different memory properties and execution paths. To fully leverage the speed up offered by these new architectures, we propose to develop two tools: A hardware adapter for restructuring existing codes as hybrid codes for specific hardware, and an optimization tool to determine optimal parameter settings as well as processor and memory allocations for specific problem instances. These tools will enable CFD code performance improvements without significant rewriting of the code itself.