The main scientific and technical problems of using hybrid HPC clusters in materials science
Abstract
The article discusses the use of hybrid HPC clusters for the execution of software designed to calculate the electronic structure and atomic scale materials modeling. Modern software systems, which are designed to solve the problems of materials science, use the capabilities of various hardware computing accelerators to increase productivity. The use of such computing technologies requires the adaptation of application program code to hybrid computing architectures, which include classic central processing units (CPUs) and specialized graphics accelerators (GPUs).The use of large computing hybrid systems requires the development of methods for ensuring the workloading of such computing systems that will allow efficient use of computing resources and avoid equipment downtime.First of all, these methods should allow parallel execution of user applications using computational accelerators. However, in practice, software environments designed to solve application problems cannot be deployed in the same computing environment due to software incompatibility. In order to overcome this limitation and ensure the parallel execution of diverse types of materials science tasks, the creation of individual task execution environments based on virtualization technologies and cloud technologies.The continuation of virtualization technologies and the provision of cloud services is the construction of digital platforms. The article proposes the use of a digital platform for hosting scientific materials science services that provide calculations using various application software systems. Digital platforms make it possible to provide a unified user interface to scientific materials science services. The platform provides opportunities for finding the necessary scientific services, transferring source data and results between users, the platform and hybrid high-performance clusters.
- Publication:
-
Materials of Electronics Engineering
- Pub Date:
- February 2020
- DOI:
- 10.17073/1609-3577-2019-4-262-267
- Bibcode:
- 2020MEE....22..262V