Estimation of the workload of a hybrid computing cluster when performing modeling tasks in materials science

Abstract

The article is devoted to methods of calculation and evaluation of the efficiency of functioning of hybrid computing systems. Material science software systems demonstrate maximum efficiency when operating on hybrid computing systems when using graphics accelerators for calculations. Examples include the VASP (The Vienna Ab initio Simulation Package) and Quantum ESPRESSO software systems. These software systems are most efficient when using monopolistic computing resources: RAM, CPU, GPU.When operating a hybrid high-performance cluster, the problem arises of resource management and their division between a group of users. Technologies need to be developed that ensure the allocation of resources to materials science applications for different users and research teams. The modern approach to organizing the computing process is the use of virtualization and cloud technologies. Cloud technologies enable the provision of SaaS and PaaS services to users. It is advisable to provide scientific teams with applied materials science systems as cloud services.Such diverse approaches, when applied in a single computer complex, require the development of methods for optimizing the load on the resources of a high-performance complex, assessing the efficiency of using its computational capabilities, and developing methods for improving user programs.Determining the quality of the complex loading is an important task when providing high-performance computing services to research teams performing interdisciplinary research in various fields of science and technology. The article proposes a method for calculating the value of the load value using the peak performance values of the complex. The results and performance quality of high performance computing cloud scientific services are analyzed using a Roofline model.