The π2.0 Supercomputer, designed and built by Inspur for Shanghai Jiao Tong University (SJTU), recently became operational. The system, with a peak performance of 2099.2 TFLOPS, ranks 368th on the lasted 53rd Top500 list.
SJTU has been committed to the construction of and investment in supercomputers, and its π1.0 system, which was listed in the top 500 list in June 2013, was also built by Inspur. π1.0 has been serving the whole university and provides computing support to the research in science, engineering, agriculture, biology, medicine, etc. Empowered by the system, the university has 72 papers published. In March 2018, Science, a world-renowned journal, featured the latest achievements of research headed by Prof. Zhao Liping’s team at SJTU’s School of Life Sciences and Biotechnology.
π1.0 (left) and π2.0 (right) supercomputer systems
However, with the number of users growing, in-depth optimization of software, and upgrading of hardware architecture, the π1.0 supercomputer system had been unable to meet existing needs—the overall platform was in urgent need of upgrading.
For the purpose of meeting the high performance computing needs across SJTU, Inspur engineer team determined which applications had consumed most of the resources of the π1.0 supercomputer system to date. It was found that materials chemistry, computational physics, life sciences, simulations for CAE, and independently developed software required the most system resources. Such operations usually have better parallel efficiency and scalability, and are suited to being run in a CPU-based distributed memory computing platform.
To cater to these scenarios and operational needs, Inspur designed a computing cluster with Intel’s latest Cascade Lake 6248 CPU. The cluster, with a total of 656 two-way nodes and 26,240 cpu cores, adopts Intel Omni-Path 1152 port core switch to provide a 100Gbps high-speed network, meeting the needs for full-wire speed, congestion-free communications during computing and storage. Inspur also built an all-flash distributed storage system, which enables 70Gbps or faster aggregate reads/writes with just five storage nodes. The system can undertake computing tasks in a temporary computing or production test environment, and concurrently process a massive number of small files.
According to Lin Xinhua, director of HPC of SJTU, the single-node computing power of π2.0 is 4.5 times higher than that of π1.0. The results of tests with examples in astronomy, oceanography, and materials genomics showed that computing performance was improved by 2 to 5 times on π2.0. Meanwhile, Lin Xinhua disclosed that π supercomputer is making the transition to a large computing platform, which will not only provide computing resources, but also offer custom development and optimization services according to user needs. After the upgrading, the system will continue to help lecturers and students in the university improve scientific research efficiency and solve more scientific conundrums.