Research On Technologies For Service Performance Optimization In Data Center Network

Data Center Network(DCN)generally becomes the core and basic infrastructure for cloud services and has attracted great attention by more and more researchers and people from industries.The centrally processing need for big data problems turns the traditional south-to-north traffic pattern to east-to-west pattern in data center networks.Traffic chanracteristics in DCN are now facing great transformation.At the same time,the performance objective of traffic in Data Center Backbone Networks(DCBN)becomes diversified.However,cloud service providers care more about operation cost and they mainly focus on the high utilization of bandwidth in DCBN.It is difficult to guarantee the service performance requirements in DCBN.Moreover,as the number of new services and users keeps increasing,the traffic in the core links of DCN becomes overwhelming,as well as the bandwidth competition becomes more fierce.All these new changes and trends in DCN bring huge challenges to service performance.It is urgent to optimize the service performance in data center networks under such new situation.Based on systematic summary of research background and the existing work,this dissertation focuses on optimization of services performance in data center networks.On one hand,in order to improve the performance of traffic transfer directly,the new techniques for coflow scheduling in DCN and aggregate traffic transfer in backbone network are carefully designed.On the other hand,virtual machine placement and migration method are investigated to optimize the network traffic distribution,which helps improving service performance undirectly.The main contributions of this dissertation are summarized as follows:1)To cope with coflow scheduling which is existed only in DCN,a deep reinforcement learning algorithm is prioposed.With the traffic characteristic changes over time,this proposed algorithm translates the coflow scheduling problem with bandwidth constraint into a continuous learning process and achives coflow transfer time minimization.By learning the previous decisions,the best scheduling is obtained.Combined with back filling and limited multiplexing mechanisms to guarantee work conserving and starvation free,the average coflow completion time is further decreased.Simulation results show that,under different network load,compared with other scheduling mechanisms,the average coflow completion time is reduced.Especially when the network load is heavy,the proposed mechanism achieves about 50% performance improvement than the state-of-the-art scheduling mechanism.Furthermore,the proposed mechanism is adaptive and achieves good performance when the traffic character changes.2)To cope with the aggregate traffic determined by performace requirements in DCBN,a NUM model is proposed.Researchers have found that traffic in data center backbone networks can be catogrized into several classes.Performance improvement has been an increasing demand for multiclass services in DCBNs recently.Unlike the simple multi-path routing schemes to balance the load and basic Network Utilization Maximization(NUM)model to maximize the whole network utility,a new NUM-based framework named Wide-Sense Circuit Path(WSCP)is proposed to optimize rate and aggregate bandwidth allocation for DCBN in the paper,with constraints on bandwidth resource and multiple performance requirements that include packet delay,loss,and traffic throughput.Because of the non-convexity of some constraint functions,the formulated NUM problem is non-convex,which is challenging to solve.To deal with such a problem,we propose sequential parametric convex approximation(SPCA)to transform it into several convex sub-problems.Then we obtain optimal solution of the formulated problem by solving the convex ones instead,using a well-known available tool named CVX.We validate WSCP via simulations on the topology of Google's data center backbone network.Numerical results show that WSCP guarantees that services belonging to different classes strictly meet their performance requirements by optimal rate and aggregate bandwidth allocation,yet runs 1.5 times faster than the other two state-of-art algorithms3)To cope with the situation that traffic brought in by VM pairs that have heavy communication with each other occupys too much bandwidth on core links in DCN,a solution towards traffic minimization on core links is proposed.In cloud data center networks,virtual machine placement can be investigated to change the traffic distribution,decrease traffic on the core links and tackle traffic scalability issue.By taking advantage of data center network architecture,we formulate the traffic scalability issue as a combinatorial optimization model of online virtual machine placement with multi-dimensional resource constraints.By leveraging Markov approximation technique,an approximation of the optimum is obtained.Theoretical analysis shows that the gap between the optimum and the approximation result is less than a tiny constant.Performance evaluation demonstrates that the proposed method decreases traffic on the core links in DCNs and achieves significant traffic scalability improvement over two common heuristics.4)To cope with the long time and large volume of data tranfering on core links of the network when a virtual machine migrates from one machine to another online,a fast live virtual machine migration method is proposed based on rack level memory page reduction.Virtual machine migration may cause too much data transfer in the core network and degrade the service performance.Traditionally this problem is solved by memory page retransmission avoidance mechanism,which uses a hash table to detect the same memory page that need to be transferd.The uneven distribution of link list length makes long time lookup and cause more data transfer in the end.In this dissertation,hash table is replaced by Bloom filter to make the link list length more even on everage,then a pruned algorithm is proposed to reduce the length of each link list and accelerate the lookup as much as possible.A rack level memory page retransmission avoidance mechanism is then constructed,which makes the same memory page transmit only once between source and destination racks,thus greatly reduces the number of transmitted memory pages.Experiment results show that the proposed method transfers less data on core links,takes less time to finish migration,and decreases the degradation of network applications caused by migration.