Scalar tuning of a fluid solver using compact scheme for a supercomputer with a distributed memory architecture

The scalar tuning of a compressible fluid solver for a supercomputer with a distributed memory architecture is conducted. We use the K computer which is one of the peta-scale supercomputers recently developed in Japan. A computational code "LANS3D" and its high-order compact differencing option are tuned. The original version of the code achieves approximately 4.5% of full performance of CPU for the simple test case. Scalar tuning based on combining do-loops works well, and the tuned code attains about 10% of full performance for the same case. The reasons are the improvement in the use of the cache, the suppression of the data transfer, and the efficient use of the data that once transferred to the cache from the memory that results in hiding the low speed of data transfer. The tuned code becomes twice faster than the original one in the wall-clock time and enables us to perform over-160-case parametric study about airfoil flow computation by large-eddy simulations with high-order accurate and high resolution numerical scheme.