Research On High Performance Network Structure And Fault-tolerant Algorithm For Novel Shared Storage Interconnection

The development of cloud computing and big data keeps the data center scale expanding,and requires that many kinds of resources including computation,storage and network,be fully utilized.In terms of memory or storage resources,the development of new storage technology is able to greatly improve the memory or storage performance,but the new storage devices are relatively expensive.This paper aims to realize a new way to staticly share storage resources of server nodes in a rack by realizing a model of stable and reliable high-speed interconnection newotk.This paper introduces the development status of new storage technology,high-speed interconnection structure and resource sharing technology.Then the key interconnection network structure and the fault-tolerant mechanism of the network are both designed.The main work of this paper is as follows.(1)The development of existing new storage technology and high-speed interconnection structure is reviewed.The trend of Rack-Scale Design and resource sharing technology in data center is introduced.(2)The design of interconnection network named RSI in racks under the the constraints of the interconnect demand and port number.RSI-FR(DF)outperforms about 12.8% than RSI-BR in throughput bandwidth when the network contains less than 256 nodes;when it comes to 512 nodes or more,RSI-FR cannot interconnect so much nodes like BR,HR and SR.The saturation throughput bandwidth of RSI-BR is higher than RSI-HR and RSI-SR at 41% and 12.3% respectively when there are 1024 nodes.At the same time,the remote access throughput bandwidth of BR can reach about 310 Gbps and the remote access delay is about 0.5us,which satisfies the bandwidth and latency requirements of the node's access to the remote new storage resources.(3)A low cost adaptive routing mechanism suitable for RSI network is designed to realize load balancing and to avoid deadlock with less VCs.The Saturation throughput bandwidth has been improved from 86 Gbps to 216 Gbps in hot spot cases.(4)An intelligent routing table updating algorithm is designed to realize the fault tolerance mechanism and to maintain the availability of network nodes with little performance degration.The test results show that when some node failures,the throughput bandwidth of super-nodes containing the fault nodes can be reduced by about 3% while other super nodes are not affected;when a whole super node fails,the performance of other normal super nodes is almost unaffected.