An Adaptive Routing Method for Interconnected Network Systems

Abstract

The invention discloses a novel Dragonfly interconnected network system. The system divides the routers in the system into n router groups, and the n router groups are fully connected, and each router group contains one or more router groups. The routers in the group are fully connected. Based on the adaptive routing method of the new Dragonfly interconnected network system, the present invention on the other hand labels the routers in the Dragonfly interconnected network system, and adopts a specific label-minus-direction priority algorithm or a label-increase direction priority algorithm, which does not need to be used. The virtual channel can realize partial adaptive routing without deadlock, and ensure that the communication step size between any processor nodes does not exceed 5 steps. The completely self-adaptive routing method proposed in another aspect of the present invention utilizes the deadlock-free feature of the partial self-adaptive routing method of the present invention and is supplemented by a novel flow control mechanism to realize deadlock-free completely self-adaptive routing.

Description

technical field [0001] The invention relates to the technical field of distributed computing, in particular to an interconnection network system and an adaptive routing method. Background technique [0002] The interconnection network usually realizes synchronization and communication between different processing nodes, and is an important part of connecting processors, memory, and I / O devices in a multiprocessor system. The interconnection network is an important aspect that affects the performance and scalability of the overall multiprocessor system. How to provide low-latency and high-efficiency communication between processing nodes on the basis of existing technologies has become a key problem to be solved urgently in the interconnection network. [0003] The development of microprocessor technology has rapidly increased the computing power of a single processing node, which also puts forward higher requirements for the performance of the interconnection network. In f...

AI Technical Summary

Problems solved by technology

The Blackwidow Clos network used high-end routers for the first time, but the network cost was high and the network diameter was large, which had a certain impact on system performance; the Cascade Dragonfly network also used high-end routers, which featured high scalability and low cost. And it can provide a smaller network diameter, but it needs to use 4 virtual channels to avoid deadlock, which increases the complexity of router design, increases message routing delay, and limits system performance

Images