NCKU Team Develops Next-Generation Network Topology Surpassing Supercomputer Technologies in US, China, and Japan
Written & edited by News Center. Image credit to News Center.
Chair Professor Chi-Chuan Hwang (the second from right) and his researching crew members
After a decade of hard work, the team led by Chair Professor Chi-Chuan Hwang at the Department of Engineering Science of NCKU has developed a high-efficiency network topology architecture called “Equality topology”. They have not only successfully obtained patents in the EU, the US, Japan, China, South Korea, and the ROC but also published their achievement in the international journal, The Journal of Supercomputing, titled “Performance evaluation of multi-exaflops machines using Equality network topology”. This network topology can design efficient supercomputers with over 10E of computing power and is a groundbreaking key core technology in the competition among the US, China, the EU, and Japan in creating 1E (the computing power of approximately 10 million desktop computers)-scale super computers.
Network topology architectures are an extremely vital part of the design and establishment of super computers and determine the computing performance of the overall architecture. Chair Professor Chi-Chuan Hwang at the Department of Engineering Science of NCKU likened supercomputer servers and switches to manufacturers and distributors; if the distributors face poor transportation conditions, the products cannot be successfully sold to make profit no matter how many products are produced. Thus, the efficiency of the switches determines whether the total computer power of the entire supercomputer can be smoothly upgraded.
"Equality network topology" can design efficient supercomputers with over 10E of computing power
“Network topology is like the road system plans of an entire country, and in terms of the traffic rules, our network topology offers greater convenience and better flow efficiency than other topologies,” said Professor Hwang. The concept of the Equality topology developed by the NCKU team lies in distributing nodes within a circle; each node is a switch. The distinguishing feature of the Equality topology is that the external world which each node sees is exactly the same; everyone begins from an “equal” starting point. In other words, Equality topology has high symmetry and facilitates the establishment and reading of traffic route tables.
In the fat tree topology, which currently occupies over half of the market, switches must search for connecting circuits between the servers of different tiers in order to engage in communication. From the establishment and verification of small systems to multi-exaflops supercomputers or data centers, the simulation efficiency of the Equality topology surpasses that of other network topologies, including the fat tree, 3D-torus, and 5D-torus network topologies commonly used on the market. The Equality topology also offers high scalability, able to connect over a million computing nodes. This technology can be used to establish zetta (1000E)-scale supercomputers and is the most important core weapon in the future competition among high-tech R&D and design industries around the world.
Professor Hwang led his students to HPC-AI competitions and won the prize every year
In 2002, Professor Hwang, who was vice president of the Office of Academic Affairs at NCKU at the time, was conducting a transnational project funded by the National Science Council to work with the IBM Thomas J. Watson Research Center in the US to build the third generation of their Blue Gene supercomputer. Although the project was ultimately terminated and no supercomputers were successfully purchased, Professor Hwang met Professor Yue-fan Deng, who was teaching at the Stage University of New York at Stony Brook at the time, through this project. Professor Deng had hit a wall regarding switch connections, and ultimately, it was the NCKU team that found a solution, which was also the turning point in their opening a supercomputer business.
However, there is a high threshold for entering the supercomputer business, and making it a success was no easy matter. Professor Hwang pointed out that throughout the US, China, and Japan, probably no more than a thousand people understand the core technologies behind these network topologies. Thus, sending their paper for review took Professor Hwang’s team a fair amount of time and effort. Furthermore, supercomputer equipment can easily cost tens of millions of dollars, not to mention the yearly patent maintenance fees, paper publication fees, and labor costs, all of which are a heavy burden for a university team. The NCKU team spent 7 to 8 years producing this paper. Regarding all of their hard work, Professor Hwang remarked, “I never wanted to give up because I know this is a significant breakthrough that has the chance to change the history of human technology.”
Supercomputers are a strategic industry in national scientific research. In 2016, the US began banning supercomputer-related exports to China, and China followed suit by regulating the outflow of their core supercomputer technologies. Both sides consider supercomputers to be a competitive high-tech industry because fields such as national defense, geology, meteorology, oil exploration, automobile, aviation, space, chemicals, pharmaceuticals, and artificial intelligence all require supercomputers for calculations and R&D. Professor Hwang pointed out that the global supercomputer market is currently about USD 20 billion a year, whereas the cloud center market is approximately USD 200 billion. The NCKU team has already signed a contract with the Fujitsu Corporation in Japan, and they plan to conduct a real machine test of their Weather Bureau Research System this March.