Network Load Balancing System Based On Lightweight In-band Telemetry

Nowadays,with the increasing application of big data and cloud computing technologies,data center networks have also become the support of cloud computing infrastructure and big data processing.Data center networks are constantly being adjusted and improved to meet the needs of current network development.This master thesis mainly studies the problem of load balancing in data center networks,and designs a network load system based on lightweight in-band telemetry for the entire network,including a network telemetry framework based on in-band telemetry and source routing,and a routing and load balancing framework based on the TopK algorithm.(1)This thesis take use of the hierarchical and regionalized characteristics of the fat tree data center network and regard the network topology as a directed graph.The network telemetry framework divide the topology into multiple Euler diagrams,and construct Euler loops in each Euler diagram which compose the whole network traversal paths.For fat tree topologies of any size,the goal of non-overlapping can be achieved,which greatly reduces the waste during information collection.At the same time,the path planning algorithm can also be applied to re-plan the path in the case of system failure,and only the links in the failure Euler diagrams need to be re-planned.It avoids the re-planning of the entire network and reduces the impact of the failure on the system,also improve the processing and operating efficiency of the system when some links fail.(2)This thesis proposed a routing and load balancing framework based on TOPK algorithm.The author select K paths for data flow to forward instead of one optimal path.Through multi-threading and distributed deployment,routing and load balancing are realized,and the efficiency and timeliness of routing are guaranteed.This thesis uses large-scale networking simulation and small-scale simulation environments to test the timeliness and operational effects of each algorithm proposed in this thesis.The author conducted simulation experiments on load balancing algorithms for different network traffic model.Through comparison of experimental results,it is proved that the architecture and strategy proposed in this thesis works well and perform better than ECMP and HULA.