| Abstract: |
|
In this work we present AWP-ODC-OS, an end-to-end optimization of AWP-ODC targeting homogeneous, manycore supercomputers. AWP-ODC is an established software package simulating seismic wave propagation using a staggered finite difference scheme which is fourth order accurate in space and second order in time. Recent production simulations, e.g. using the software for the computation of seismic hazard maps, primarily relied on GPU accelerated supercomputers. In contrast, our work gives a comprehensive overview of the required steps to achieve near-optimal performance on the Intel Xeon Phi x200 processor, and compares our competitive performance results to the most recent GPU versions. At the level of a single vector operation, we apply the vector folding technique to AWP-ODC-OS, yielding a 1.6× performance increase over traditional vectorization. Further, we present a novel strategy utilizing both DDR4 RAM and High Bandwidth Memory, increasing the maximum problem size operating at maximum performance by 26%. The presented shared and distributed parallelization carefully schedules work to the cores and ensures overlapping communication and computation. We conclude with a detailed study of AWP-ODC-OS’s full-application performance on the Intel Xeon Phi x200 processor, achieving up to 97% of the most recent P100 GPU generation’s performance. Additionally, our weak scaling study on up to 9000 nodes of the supercomputer Cori Phase 2 achieves a parallel efficiency of greater than 91%, equivalent to the performance of over twenty thousand K20X GPUs. |
|
;){kind=link}