Performance Evaluation of Large Scale Electron Dynamics Simulation under Many-core Cluster based on Knights Landing

Abstract

We have been developing an advanced scientific code called "ARTED" for an electron dynamics simulation using the first-order computation of materials to be ported to various large-scale parallel systems including the "K" Computer, which was previously Japan's fastest supercomputer. In this paper, the implementation and performance evaluation of the ARTED code used in Intel's latest manycore processor, the Knights Landing (KNL) stand-alone cluster, are described based on past research on porting the code to the Knights Corner (KNC) accelerator. Our target system is Oakforest-PACS, which is currently the fastest supercomputer in Japan. For performance tuning on KNL, the largest issue is how to utilize multiple levels of parallelism, such as the instruction level (512-bit SIMD instruction), hardware thread (4 threads/core), and large number of cores. We focus on the dominant computation part of the code, where 25 points of a 3D stencil computation are required. We successfully optimize this part to achieve 758.4 GFLOPS per node, which corresponds to 24.8% of the theoretical peak on the node of Oakforest-PACS using an Intel Xeon Phi 7250 (3046 GFLOPS peak). It is also shown that the KNL sustained performance is better than that of the two KNC accelerator cards. The entire ARTED code implies other time step computing, and was designed for a large-scale parallel execution using MPI, whereas single-node parallelization is achieved using OpenMP. We finally evaluate the entire parallel execution performance with up to 128 nodes.

Performance Improvement Options of Scientific Applications on XeonPhi KNL Architectures

“…Recently, with the emergence of Intel's KNL architecture, HPC researchers have endeavored to fine tune their applications or do a performance analysis study of applications in order to ease parallelism on a KNL machine. For instances, Barnes et al (2016) have analysed National Energy Research Scientific Computing Cente (NERSC) workloads on KNL; Afanasyev and Voevodin (2017) and Liu et al (2017) have manifested that executing graph algorithms in KNL is better than GPUs -in these works, the authors have investigated the cases where manycore features and scalability features of applications could be improved in KNL; Hirokawa et al (2018) have studied the PI options while porting and executing their Ab-initio real-time electron dynamics (ARTED) code on KNL machines. This paper has studied the Intel KNL architecture and the PI techniques for HPC applications on KNL.…”

Performance improvement options of scientific applications on XeonPhi KNL architectures

“…Recently, with the emergence of Intel's KNL architecture, HPC researchers have endeavored to fine tune their applications or do a performance analysis study of applications in order to ease parallelism on a KNL machine. For instances, Barnes et al (2016) have analysed National Energy Research Scientific Computing Cente (NERSC) workloads on KNL; Afanasyev and Voevodin (2017) and Liu et al (2017) have manifested that executing graph algorithms in KNL is better than GPUs -in these works, the authors have investigated the cases where manycore features and scalability features of applications could be improved in KNL; Hirokawa et al (2018) have studied the PI options while porting and executing their Ab-initio real-time electron dynamics (ARTED) code on KNL machines. This paper has studied the Intel KNL architecture and the PI techniques for HPC applications on KNL.…”

EigenKernel

“…First, we show the performance data of various GEP solvers on Oakforest-PACS obtained using EigenKernel. Such data will be of interest on its own since Oakforest-PACS is a new machine and few performance results of dense matrix solvers on it have been reported; stencilbased application [11] and communication-avoiding iterative solver for a sparse linear system [12] were evaluated on Oakforest-PACS, but their characteristics are totally different from those of dense matrix solvers such as GEP solvers. Furthermore, we point out that one of the solvers has a severe scalability problem and investigate the cause of it with the help of the detailed performance data output by EigenKernel.…”