Research On Flow Scheduling Of Data Center Network Based On SDN

A conventional Data Center Network(DCN)mainly has the following disadvantages.First,the conventional tree-based topology has limited capacity.Second,Equal-Cost Multipath Protocol(ECMP)is not able to take full advantages of multipath diversity.Third,it not able to use network resource effectively because lack of intelligent management.We research the flow scheduling problem for DCN.We use Software Defined Network(SDN)technology to managed DCN intelligently,and adopt Fat-Tree topology that has an oversubscription of 1:1,and solve the low link utilization problem caused by ECMP through realizing an efficient flow scheduling mechanism.When we research Hedera,we found the flowing problem in Hedera.Firstly,Simulated Annealing(SA)algorithm has the flow conflict problem because it does not consider the current network link bandwidth resources.Secondly,SA adopt the method of one-to-one mapping from destination to core switch,so it don't schedule the flows that forward from aggregation switch.Finally,Global First Fit(GFF)algorithm is easy lead to the irrational distribution of bandwidth resources due to not consider the change of flow demand.To solve these three problems,we propose an adaptive on-demand flow scheduling mechanism.The mechanism is composed of two main modules.There are the filter strategy based on the changing of demand and the scheduling paths search algorithm based on simulated annealing genetic algorithm.The former module filters the flows that is newly arrived or of which new demands not meted by the assigned shcheuling path,and then adds these flows that not meet demand into to the scheduling flow set.The latter module adopts the method of mapping a destination host to the highetst level swith and calculates scheduling paths for the scheduling flows based on flow demand and the available link bandwidth resources.We implement the mechanism by POX and Mininet and compare it with ECMP,GFF and SA in experiments.The results show that the mechanism we proposed outperform ECMP,GFF,SA and Ashman-Prob Fit in Stride,Staggered and Random communication patterns,especially,it's lower than Ashman-Best Fit in Sride(4)and Sride(8)and lower than SA and Ashman-Best Fit in Randombij.