Reasearch And Implementation Of Data Center Network Traffic Scheduling

The data center is an essential infrastructure for service providers to host various IT businesses.In recent years,the rise of the cloud computing business has accelerated the widespread deployment of the cloud data center,and the traffic composition of a data center network has also undergone structural changes.Intra data center traffic accounts for the majority of the total traffic,and the scheduling of these traffic has become the key to the optimization of data center network performance.The data center traffic scheduling problem includes two subproblems:load balancing and priority scheduling.The problem of load balancing focuses on how to balance the load among multiple equivalent paths so that the whole network can better cope with traffic growth.The priority scheduling problem focuses on how to give different priorities to different traffic demands to better meet their network performance requirements.In this paper,two load balancing schemes LC,HB and two priority scheduling schemes PB,CB are proposed.Based on the functions of existing switching devices and new programmable switching devices,LC and HB implement the distributed congestion aware load balancing mechanism.PB and CB are different from the existing priority scheduling schemes,which use packet size and congestion window size instead of flow size as the basis of dividing flow priority.The simulation experiment also verifies the effectiveness of the traffic scheduling synthesis scheme.The innovation of this paper can be summarized as follows:(1)Combining load balancing with priority scheduling,a comprehensive traffic scheduling scheme is designed.This comprehensive scheme not only considers the design difficulties under the two subproblems but also considers the problems such as inaccurate load information that may occur when the two are combined,so it has enlightenment for the follow-up research.(2)In the aspect of load balancing,LC and HB are designed under the existing switching equipment(low cost)and new switching equipment(sufficient cost),respectively.In the LC scheme,probing and piggybacking are used to sense the path load.The load and the transmission rate are used as the primary and secondary basis of path selection to avoid herd effect.In the HB scheme,only piggybacking is used to detect the queue length,which can avoid the extra overhead of probing while ensuring the visibility of the network.(3)In the aspect of priority scheduling,we should avoid using the flow size as the basis of priority(FB scheme)but use the packet size(PB scheme)and congestion window size(CB scheme)as the basis.Without modifying the transmission protocol,the phenomenon of large flow starvation in FB scheme is avoided,and the impact on large flows is weakened while the performance of small flows is improved.