Performance test of Taurus HPC system
Hailong Zhang, Jun Nie · 2016
High performance computing system has become the strong support of scientific research due to the good cost performance ratio and scalability after pure science and experimental science, the performance tests help to find out the performance bottlenecks of HPC systems. How to evaluate the performance of a HPC system is the main purpose of this paper. A high-performance computing system with 16 CPU+GPU compute nodes has been constructed. HPL benchmark was used to evaluate the performance of CPU, GPU under different matrix scale Ns, computational nodes, block size NB of matrix etc, and an empirical formula of Ns value calculation was presented through lots of testings. A comprehensive storage I/O performance was tested with IOZone which is a professional file system's standard testing tool. The performance of CPU cluster reaches 93.5% of the system theoretical value and GPU cluster gets 80% of the theoretical value.