A Link Fault Detection Method in HPC Indirect Network Environment

Abstract

The invention provides a link fault detection method in a HPC indirect network environment. According to the link fault detection method based on link delay information measurement applied in the HPCindirect network environment, the link fault is detected by detection of anomaly of link delay information; and thus, a fault link in a network can be precisely determined within a relatively short time. The link fault detection method comprises the steps of: (a), querying HPC interconnection network routing information, so that the link composition of communication paths between nodes is obtained; (b), in combination with the link composition of each communication path, determining a key communication path set needing to perform delay measurement; (c), measuring delay information of a key path in parallel, and, on the basis of the information, solving delay information of all links in the whole network; and (d), judging whether the link fails or not according to the link delay information, and solving the link delay expected value in the network, wherein the delayed link, which has relatively large deviation with the value, is the fault link.

Description

Technical field: [0001] The present invention relates to a link fault detection method, and more specifically, to a link fault detection method in an indirect network environment of a High Performance Computer (HPC for short). Background technique: [0002] High-performance computing refers to the process of using certain technologies to aggregate the computing power of a large number of processing units to solve complex problems. High-performance computing has gradually become an important means to solve major challenges in national economic construction, social development, technological innovation, and national security, and it is a strategic high ground for competition among countries around the world. The scale of high-performance computing continues to increase, and it is now moving towards the E-level scale. The scale of computing of such a volume corresponds to a very large scale of computers. Take the Sunway TaihuLight high-performance computer, which ranks first in...

AI Technical Summary

Problems solved by technology

However, in the HPC interconnection network environment, due to the consideration of communication efficiency, there is often no sufficient communication protocol support, making these methods often cannot be directly used to detect link failures in large-scale high-performance computer interconnection networks, and because HPC interconnection The network scale is often very large, and the time overhead of the existing link fault detection method is very huge under this network scale

Images