Communications in Computer and Information Science
Modern Elbrus-4S and Elbrus-8S processors provide a level of floating-point performance close to that of widespread x86_64 CPUs that are predominantly used in high-performance computing (HPC). The uniqueness of the software ecosystem of Elbrus processors requires special attention in the case of their deployment for execution of mainstream computational codes. In this paper, we consider the performance of one widely used code for computational materials science (VASP), as well as FFT libraries. The results for the Elbrus processors are embedded into the context of performance of modern x86_64 CPUs.
The paper describes the first experience of practical deployment of the hybrid supercomputer Desmos at the Joint Institute for High Temperatures of the Russian Academy of Sciences (JIHT RAS). We consider job scheduling statistics, energy efficiency, case studies of GPU acceleration efficiency and benchmarks of the distributed storage with a parallel file system.
In the article a technology is considered which aims at creating architecture-independent parallel programmes based on the data-driven functional paradigm. A proposed toolkit provides the translation, execution, debugging, optimisation and verification of programmes. A programme in a data-driven functional parallel language is translated into the data-flow graph (which describes the data dependencies of an implemented algorithm) of the programme. On the basis of this representation, the control-flow graph (which defines the organisation of computations) is generated. Both graphs allow to carry out various optimising transformations. The resulting data-flow graph is also used for the formal verification of the programme. A computation process is considered as a cooperation of the control-flow graph and the data-flow graph. The execution of data-driven functional parallel programmes is carried out by a special interpreter (event machine), which consist of a number of event processors controlled by a special manager.