Research On Data Center Network Resource Management And Performance Optimization

Data centers are the key infrastructure of modern Internet and cloud computing.Nowadays,the amount and scale of data centers are a significant symbol of the information technology development level of a country and have become a strategy competition point of all nations.It's a long time since Data centers have emerged.However,the research on data centers is just in its nascent and burgeoning stage.Especially in recent years,with the quick development of Internet and cloud computing,the composition,structure,function,scale and application mode of data center networks have been profound changed.Network resources are the critical resources in data centers.And network resource management and performance optimization are two basic aspects for keeping data centers working effectively.So it's of significant theoretical and practical importance for us to thorough research on them,which can help us to understand data center networks' work mechanism,improve their performance,and save investment and cost.Automatic address configuration,transport control protocol and traffic management technology,which have key effects on network management and performance optimization,are three of the most important aspects in modern data center networks.In this paper,we focus and research on these three aspects.We start from the basic characteristics research on data center networks,and aim at a network performance optimization.The main work and contributions are as follows:1.Comprehensive analysis and study on the characteristics of modern data center networks.Since network characteristic research is the basis of network resource management and performance optimization.In this paper,we study a couple of basic characteristics of modern data center networks from several important aspects,including structure characteristics,application properties,traffic characteristics and virtualization features and get some preliminary conclusions,including: ?1 In the structure characteristic aspect,basic network structures of modern data center networks are transforming from traditional tree-like structures to fat-tree and non-tree structures.?2 In the application mode aspect,modern data center networks are changing from private mode to cloud computing fashion.Network resources are shared dynamicly by multitude tenancies and applications.?3 In the traffic characteristic aspect,cluster communication is the typical communication pattern in data center networks.The traffic in data center networks is not well-distributed at all.It exhibits a notable locality and dynamic characteristic.?4 In the network virtualization aspect,virtualization has been widely adopted by current data center networks.And VM has become the basic unit of resource allocation and scheduling.To supporting online VM migration and dynamic resource management,data center network should be a large two layer network and support dynamic VLAN division.Those new characteristics of modern data center networks above propose challenges to their resource management and performance optimization,such as automatic address configuration,congestion control and traffic scheduling,and thus new technologies is needed to address them.2.Research on automatic address configuration of topology-dependent data center networks.The addresses of many newly proposed data center network structures are topology-dependent,that's the addresses of nodes are related to their locations and network typologies.The key issue of topology-dependent address configuration is logical address to physical address mapping,which can be solved by graph isomorphism theory intuitively.However,graph isomorphism is hard.Recently proposed solutions all base on graph isomorphism theory,making them still suffer from some shortages,such as efficiency and error tolerance problem.In this paper,we first propose a generic and error-tolerant automatic address configuration method GA-MAP for generic data center networks,and then a quick automatic address configuration method TREE-CONF was proposed for a kind of special but widely used data center networks—tree-like data center networks.In GA-MAP,we first transform the address mapping problem into an optimization problem and solve it by an improvement genetic algorithm.Our algorithm is effective and error-tolerant in nature.It even can automatically configure the nodes involved in malfunctions with right addresses.TREE-CONF is a special automatic address mapping algorithm for tree-like data center networks.By leveraging the special structure characteristics of tree-like data center networks,it can usually finish the address configuration process more effective than a general isomorphic mapping method.For a network with millions of nodes,it can finish the mapping process within several seconds.3.Proposed a deadline-aware/overtime cost aware data center network transport control protocol respectively for latency sensitive applications.Recent studies show that many applications in data center networks confront a certain soft-real-time constraints,namely deadlines,and missing these deadlines may affect user experience and thus return of investment.Thus directly using the traditional TCP in data center networks,which is deadline agnostic,may suffers from efficiency problem.To address this problem,we design a deadline-aware transport protocol DA-TCP.DA-TCP adjusts the send rate of flows according to their deadlines,and thus tries its best to satisfy the deadlines of all flows.But in some cases,we may concern more about overtime cost than overtime itself.In such case,how to allocate network resource to flows to minimize the total overtime cost becomes another problem.To address this problem,we further develop an overtime cost aware network protocol CA-TCP.It adjusts the send window of each flow by both overtime cost and extent of congestion.Our simulation shows that comparing with other protocols,CA-TCP usually has the minimal overtime cost.4.Designed two traffic balance methods respectively by flow scheduling and VM migration.The number of flows in data center is always large.Large number of flows concurrently traverse the network is apt to induce traffic unbalance,congestion,packet loss and thus performance impairment.Meanwhile,the special characteristics of data center networks,such as high bandwidth,low decay and high throughput and dominated in number by mice flows,further bring severe challenges to flow scheduling.Based on different traffic characteristics and traffic scheduling requirements,we propose a traffic balance method SFTB based on flow scheduling and another method NVMS based on online VM migration respectively.By leveraging the multipath characteristic of modern data center networks and the congestion feedback mechanism of TCP,SFTB can schedule flows to spread along different paths in RTT scale,and thus immediately avoid congestion.In contrast,NVMS balances network traffic by online VM migration according to long-term traffic distribution.However,to avoid the cost brought by large amounts of traffic computing and VM migration,it adopts a distributed increment migration manner,making it can control the trade-off between the traffic balance and migration cost.By the combination of long-term traffic redistribution and immediate flow scheduling,we can balance network traffic and improve network efficiency to the utmost.