Researchers at one of the oldest high schools in Central and Eastern Europe, Vilnius University have been using high-performance computing devices for more than ten years. Technology has developed rapidly over the last decade, and research and its applications around the world have changed accordingly. Previously implemented technologies have limited the university community’s ability to compete effectively internationally. Thus, it was imperative for the university to fundamentally upgrade its existing infrastructure.
Most of the equipment produced, when used alone, is dedicated to relatively narrow tasks. No one such device can perfectly meet all the needs of a separate organization. A growing global practice is to design pools of this equipment in a way to best meet the specific complex needs of a particular organization. The main benefit of such collections of equipment is that they allow to efficiently and quickly obtain the required tool in the face of a specific complex computational task. Besides, the specific architectural complexity and completeness of such a cluster provide speed and efficiency.
A cluster has been designed
High-performance computing systems and their components are currently manufactured by many global technology companies. In recent years, high-performance computing (HPC) clusters have been upgraded and their computing performance increased by using graphics processing units, and NVIDIA is the market leader in this segment production.
For the physical (technological) part of the solution, “Novian Technologies” (previously – BAIP) architects designed a complex high-performance computing cluster consisting of 1,728 units of non-accelerated computing cores from Intel and 20,480 units of computing cores from NVIDIA manufacturer “Tensor”.
The computing cluster will have more than 13TB of RAM. The total calculated computational efficiency of this HPC cluster will reach 361 TeraFLOP (floating-point operations per second) DP (dual precision) and 4 PetaFLOP computational efficiencies in the field of deep learning. Thus, the computational performance of the current cluster TeraFLOP will increase about 12 times.
Data storage solution
A two-tier data storage solution based on the Luster ZFS file system has been designed for data storage. The control and data module will have about 20TB of fast, SSD-based useful capacity, and all other data will have about 0.8PB of useful capacity.
Computing and data storage resources will be combined using InfiniBand EDR, 100Gb / s high-speed, low-latency networking technology. This will allow computing and data storage resources to exchange information 2 times faster than in the current VU MIF computing cluster.
For optimal use of resources and integration of the designed physical (technological) part with other HPC clusters, service management, self-service, and connection software module will be installed.
Will combine HPC nodes
The software solution will ensure the possibility to connect all high-performance computing nodes both with VU-owned and external, non-VU-owned, other high-performance nodes. The implemented service management, self-service, and integration software module will allow Vilnius University to combine the available resources HPC into one common computing resource, as well as to participate in EuroHPC and similar projects.
In parallel, a business plan, service catalog, and system accounting module are being prepared for the university. Thus, it will be possible to offer the market a sufficiently wide package of different technologies, easily adapt it to the customer’s needs, promptly prepare the required environment and then implements its accounting. It is expected that the opportunities provided by the renewed and modernized VU HPC cluster will be successfully used by both the university community and the country’s advanced businesses.
The future cluster is expected to be at least 10 times more efficient than the one currently in use. The implemented solution will undoubtedly give Vilnius University a competitive advantage among the world’s research institutions. During this project, “Novian Technologies” will implement and integrate the most modern technologies of global manufacturers. Our proposed combination of technologies from different manufacturers will allow efficient use of computing resources, prompt selection of the necessary technologies to solve specific tasks. Training in neural networks will be incomparably faster.
The project is planned to be completed in the spring of 2021.