The new supercomputer for scientific analyses and simulations has a modular design; a corresponding test system for future modular supercomputers will be launched at the Jülich Supercomputing Centre (JSC) before the end of this year. An additional data analytics module is planned for the current prototype, which has been optimised for big data analyses and special requirements of applications in the field of machine learning and deep learning as well as artificial intelligence (AI).
Such a supercomputer consists of several modules which can be combined according to the modular principle as required. The first module, the DEEP-EST cluster, was recently installed at JSC by the German supercomputer supplier MEGWARE. Two further modules designed for parallel computing and data analysis are to follow by the end of the year.
Researchers will use the system to test how well the new concept of a modular supercomputer, which has been extended over several project durations, works in practice. "The scientific codes come from the neurosciences and earth observation. Others serve the research of dynamic chemical processes or the prediction of solar storms," explains Dr. Marlene Bamberg of JSC, who has been a member of DEEP-EST's project management team for more than six months.
The hardware and software used must fulfill different functions for different scientific applications. "Today we see how our scientific users increasingly combine different simulation models. The approaches come both from the field of high performance computing and data analysis, so that we can hardly meet all requirements with a monolithic supercomputer," explains Prof. Thomas Lippert, Director of the JSC.
Jülich experts have therefore been working together with European partners in DEEP projects for eight years on a new modular supercomputer architecture that can be used more flexibly and efficiently than previous systems for scientific applications. Such a modular supercomputer is no longer monolithically constructed from identical computing units. Instead, the approach envisages several specialized building blocks that can be combined as required. Complex calculations are then carried out on the cluster module using powerful general-purpose processors. Simple program parts are taken over by the booster module, which allows efficient and fast processing of easily divisible tasks with highly parallel multi-core processors.
With the two Jülich supercomputers JUWELS and JURECA, two supercomputers have already been designed and are in operation as cluster booster systems, which are among the fastest in the world. In the DEEP-EST project, which started in July 2017, further modules have been added to the cluster booster architecture of the previous DEEP and DEEP ER projects.