A Research On Bandwidth Fragmentation In Datacenter Network Load Balancing

Current flow-based datacenter network load balancing methodsmainly focus on accommodating the bandwidth demand of the flow, whileignoring the difference between the bandwidth demand of the flow and theresidual capacity of the links. This may lead to much available bandwidthon multiple paths which cannot be merged to accommodate a single flow.Aiming at the bandwidth fragmentation problem mentioned above,this paper proposes two flow scheduling algorithm with analogy to thesolution of memory fragmentation problem in operating systems, whichare called Ashman-BestFit and Ashman-ProbFit respectively.The contribution of this work involves: a) This paper proposes andanalyzes the bandwidth fragmentation problem based on existingdatacenter flow scheduling algorithms; b) This paper proposesAshman-BestFit flow scheduling algorithm and implementing it withOpenFlow protocol; c) Aiming at the imbalanced link load that may becaused by Ashman-BestFit algorithm, this paper proposesprobability-based Ashman-ProbFit algorithm.The Mininet-based experiments show that under different datacentercommunication patterns, the methods proposed by this paper achievehigher throughput rate and bandwidth utilization compared to methodsproposed by other research works and traditional ECMP methods. Inaddition, as the ratio of the flows through the core switch increases,Ashman-ProbFit algorithm achieves higher throughput rate and bandwidthutilization than Ashman-BestFit algorithm.