Research On A Datacenter Network Traffic Scheduler Based On Dynamic Priority

With the development of information technology,the scale of data has gradually increased,and more and more applications have begun to migrate to the datacenter.While the highly interconnected datacenter infrastructure brings performance improvements to applications,these applications also bring new challenges to datacenter network.How to deal with a huge number of network flows with complex characteristics has become an urgent problem in datacenter network.By arranging the network flows in the datacenter network in an appropriate order or allocating the appropriate bandwidth to the flows,the datacenter network scheduling strategy can achieve a low average network flow completion time or meet a specific Service-Level Agreement(SLA).Most of the existing datacenter network scheduling strategies use static characteristic information such as the size and width of the network flows as the scheduling basis.However,static information has the limitations of not being applicable to all scenarios and cannot truly reflect the real transmission time.In order to solve the above problems,this paper proposes the Time Machine Scheduler(TMScheduler)based on the duration of the network flows.TMScheduler has two core components:(1)Mutable Priority Queue(MPQ)that can dynamically modify the priority;(2)Time-aware Least-Attained Service(T-LAS)that uses flow duration to make scheduling decisions.MPQ realizes the function of dynamically adjusting the priority by adding an interface for modifying the priority to the traditional priority queue.T-LAS can efficiently use duration information for scheduling,and the priority mechanism based on percentage threshold in T-LAS can be flexibly adapted to different workloads.With the support of MPQ and T-LAS,after receiving the feedback of the duration,TMScheduler can dynamically adjust the priority of network flows that were previously sent to the network but are still waiting in the queue by sending control messages.This paper has implemented and tested TMScheduler in the ns-2 simulator and testbed respectively.The test results show that TMSchduler,which considers the duration of the network flows,has an overall performance improvement of nearly 50% on the websearch and datamining workloads compared with the traditional datacenter network scheduling strategy.The performance improvement of T-LAS can reach about 80% of the optimal heuristic scheduling strategy under both light-tailed workload and heavy-tailed workload.