Benchmarks for comparison in multiple HPC applications

On the annual Supercomputing Conference (hereinafter referred to as SC), there is an unofficial theme every time. The focus of the past two years has been on Machine Learning and Deep Learning, which was preceded by large-scale data-intensive computing and the potential to use the cloud to reshape the future of supercomputing.

All of these themes have one thing in common: they are not focused on CPU processors. In fact, they are all related topics related to CPU performance improvement or ecosystem development under the X86 architecture. Recall that in the last year we saw hardware devices becoming the core theme of the conference, or when the first large-scale GPU supercomputers entered the top 500 list, but they were only the accelerator, not the core of the device.

On this year's SC17, the core processor once again became an unofficial theme. The ARM-based hardware and software ecosystem has been extensively demonstrated, and supercomputers such as Cray have demonstrated comprehensive system integration, with many benchmarks matching enough to compete with Intel's most advanced products.

The ARM-based "Isambard" supercomputer will contain 10,000 cores (using Cavium's ThunderX2 ARM processor) and will be available at the University of Bristol next year. The team responsible for developing the supercomputer has been working on ARM-based HPC. Systematic research and development work. We now see that in the ARM system based on a lot of development work, the most famous is the Mont Blanc of the Barcelona Supercomputing Center (a few years ago started using the dual Cortex-A15 ARM, now based on the Cavium ThunderX2 ARM processor).

McIntosh-Smith and his team have released some very meaningful benchmark results. The benchmark used a Cray 8-node device cluster and a 32-core ThunderX2 ARM processor and compared it to Intel Skylake and Broadwell solutions. The benchmarks for comparison across multiple HPC applications are listed below -

Overall, McIntosh-Smith says that any application that is memory-intensive can run well on ThunderX2 with significantly better performance than Skylake. However, for applications that are biased towards floating-point, Skylake is better than the wider vector operator, but ThunderX2 is comparable to the Broadwell platform. If you continue to increase the high bandwidth memory, what will happen to the test results? This will be very interesting.

See the following figure for the benchmark results:

The memory-intensive advantage is most evident in HPC applications on OpenFOAM, an open source CFD application that is often used in high-performance computing for business and scientific research. The benchmark report in the above chart shows that ThunderX2's OpenFOAM test results are better than Skylake and Broadwell.

The weather and climate simulation code also shows that the performance of memory bandwidth-intensive applications will increase the same. The above image shows the production code of the Met Office on ThunderX2 - Nemo is a marine simulation code.

"Preliminary results show that for compute-intensive applications such as GROMACS, CP2K, and VASP, the performance gap between different processors is small, while memory-bandwidth-intensive applications can clearly see the differences between different processors. This is because, although the code can benefit from the wider vector units of the X86 processor, ThunderX2 can be compensated with more cores and higher clock speeds," McIntosh-Smith said.

With the advent of the high-end Cray XC50 system, we will see more test results for supercomputers in real-world production environments based on ARM processors at the next Supercomputing Conference.

See the following figure for the Isambard project architecture:

To this day, many HPC ARM observers know that Isambard is a Cray model; but if they know that the supercomputer manufacturer has chosen a more difficult path, the ThunderX2 ARM processor is integrated with the industry-renowned Aries interconnect chip, and If you can run a full set of Cray software, many people will be surprised. In comparison, adding ThunderX2 to the CS Storm series, rather than the Aries-based product line, would be much simpler for Cray. But in McIntosh-Smith's view, this is a clear indication of Cray's firm determination to use ARM extensively in the HPC space.

McIntosh-Smith believes that different ARM options may use the same dedicated deployment method in the future. "In the future, ARM-based HPC will enhance the vector function to the same level as other CPU manufacturers. The next-generation ARM products will have a vector width comparable to that of any other vendor."

The interesting point is that these results are based on a simple optimization of the number of cores and the compilation of the basic conditions, only a few hours of fine-tuning. McIntosh-Smith said that these rapid developments today are actually the result of years of unremitting efforts on ARM-based HPC systems. After the hardware is ready, the software required by ARM in HPC applications will be faster. With the official launch of the ARM processor for supercomputers, the entire industry will usher in a new era of architecture transformation. The ARM architecture will be unveiled in the HPC field. Open a new chapter in a wide range of processor choices.