The Gyrokinetic Particle Simulation of Fusion Plasmas on Tianhe-2 Supercomputer.

We present novel optimizations of the fusion plasmas simulation code, GTC on Tianhe-2 supercomputer. The simulation exhibits excellent weak scalability up to 3072 31S1P Xeon Phi co-processors. An unprecedented up to 5.8x performance improvement is achieved for the GTC on Tianhe-2. An efficient particle exchanging algorithm is developed that simplifies the original iterative scheme to a direct implementation, which leads to a 7.9x performance improvement in terms of MPI communications on 1024 nodes of Tianhe-2. A customized particle sorting algorithm is presented that delivers a 2.0x performance improvement on the co-processor for the kernel relating to the particle computing. A smart offload algorithm that minimizes the data exchange between host and co-processor is introduced. Other optimizations like the loop fusion and vectorization are also presented.