Construction method for interconnection network of massively parallel computing system

Abstract

The invention discloses a construction method for an interconnection network of a massively parallel computing system. The OCT interconnection network is formed by 8*2k*2m nodes. Firstly, every eight nodes are connected to form one Octagon interconnection network, 2k*2m Octagon interconnection networks are obtained in total, and the nodes in each Octagon interconnection network are encoded through four-digit Johnson codes; secondly, the 2k*2m nodes which are the same in node code in each Octagon interconnection network are connected to form one Torus interconnection network, wherein each line of each Torus interconnection network has 2m nodes, each column of each Torus interconnection network has 2k nodes, and eight Torus interconnection networks are obtained in total; the nodes in each Torus interconnection network are encoded through (k+m)-digit Johnson codes, and namely the OCT interconnection network can be recorded as the OCT (k, m) interconnection network, wherein k and m are natural numbers and are parameters of the number of the nodes of the interconnection network. According to the construction method, expansion of the interconnection network is performed under the condition that node degree is kept unchanged; the OCT (k, m) interconnection network is a symmetrical and regular interconnection network, the Johnson coding method is adopted for node coding, codes of any adjacent nodes only have a one-digit difference, and therefore the routing algorithm is simple and efficient.

Description

technical field [0001] The invention belongs to the technical field of parallel computing, and in particular relates to a method for constructing an interconnection network of a large-scale parallel computing system. Background technique [0002] With the continuous development of hardware technology, especially the development of VLSI technology, large-scale multi-processor systems containing thousands of processors become possible. For example, Tianhe-1A has 7168 computing nodes, and Fujitsu's supercomputing system has more than 80,000 computing nodes. In the next few years, new applications and algorithms will drive the number of processor cores on a single chip to reach the number of nodes in large-scale supercomputing systems established in the 1980s. We are entering the era of Exascale Computing. It is estimated that in 2018, the supercomputing system will reach 1exaFLOPS (1018FLOPS), entering the era of Exascale Computing. [0003] Entering the era of exascale compu...

AI Technical Summary

Problems solved by technology

The disadvantage is that the network is not scalable

Images