Asynchronous and synchronous models of executions on Intel® Xeon Phi™ coprocessor systems for high performance of long wave radiation calculations in atmosphere models

Abstract

Long Wave Radiation Calculations are one of the most time-consuming calculations in atmosphere modeling. In this work, we explore two models for executions of these calculations on Intel® Xeon Phi™ Coprocessor Systems. In the asynchronous model, we offload the radiation calculations to the coprocessors and simultaneously execute calculations on the coprocessors along with the other atmosphere model calculations in the CPU cores. In the synchronous model, the CPU cores after offloading, wait for the results, and use the results in the same time step. We developed various techniques to complete these synchronous executions in minimal time, including loop rearrangement and low-cost interpolations. Using our experiments on an Intel Xeon Phi cluster, we show that our asynchronous execution model results in savings of many months in wall-clock execution time for multi-century climate simulations. Our synchronous execution model results in performance improvements of up to 70% in long-wave radiation calculations.