Research and development of methods for the efficient implementation of graph algorithms for modern vector architectures
The development of efficient implementations of graph algorithms is an extremely important problem in modern computer science, since graphs quite successfully model many objects of the real world from different applied areas. Due to the significant size of real-world graphs, the use of supercomputers is vital for both the acceleration of calculations and the placement of graph objects of such large sizes in memory.
The given paper describes a method based on the selection of vector-oriented abstractions of calculations and data, which made it possible to create the world’s first efficient and high-performance implementations of graph algorithms for vector architectures such as the novel NEC SX-Aurora TSUBASA vector system, Intel KNL processors and NVIDIA graphics accelerators. The developed implementations of graph algorithms demonstrate significantly higher performance in comparison with the existing analogs for both multicore CPUs and NVIDIA GPUs, as well as have significantly higher energy efficiency. On the basis of the selected abstractions, a VGL (Vector Graph Library) graph framework was developed; it extremely simplifies the nontrivial process of developing and optimizing graph algorithms for vector systems.
More information on the seminar and the link to connect via Webex are available at в Indico.