Solving non-linear chemical equations for gas cooling with GPU offloading using OpenACC

Simulation of cooling effects in a gaseous system involves solving a set of non-linear PDEs for chemical reactions iteratively. The complexity and the computational requirement of the simulation grow heavily as the number of particles and the number of cooling mechanisms is increased. Fortunately, it is now possible to offload such simulations on a variety of scalable computing platforms. In this work, we refactored an implementation of gas cooling in Gadget-2. We improved the implementation further by enabling the offloading of computation on multiple GPUs using MPI and OpenACC. Finally we performed a detailed comparative study of the code performance on x64 86 (Intel Xeon CPU), ppc64le (IBM Power9 CPU), Pascal (Nvidia P100 GPU) and Volta (Nvidia V100 GPU) architecture based computing platforms.