Research On Traffic Scheduling Problems In Multi-resource Environments Of Data Centers

As the important infrastructure for hosting various services and applications,the development of data centers covers multiple stages,including enterprise data centers,modular data centers,cloud computing data centers,and so on.The scales of morden data centers are large,and they play an even more important role.The research about the problems in data centers has attracted great attention of academia and industry.On the one hand,data flows deriving from different applications have diverse transmission requirements on the data center networks,e.g.,fast response,stable transmission rate,and deadline guarantee.On the other hand,network congestion and transmission failures are frequent in data centers,and they make the transmission requirements of flows hard to be guaranteed.While obviously improving the network performance,the vast deployment of network function devices in data centers also brings new problems and challenges.The deployment scale of these devices is on par with traditional network switching devices.The latter can only route and forward the incoming packets.However,network function devices perform more kinds of function processing based on the content of packets.The execution of network functions rely on multiple hardware resources in these devices.Incoming packets need to be sequentially processed on these resources,and different kinds of network functions consume different amounts of these resources.In such multi-resource environments,the resource contention among data flows will be more intense.Some schemes used in multi-resource environments strive to solve this problem through providing fair service for flows.However,these schemes are also inappropriate in some situations.Firstly,making scheduling decisions based on strict fairness results in waste of resources in multi-resource environments.This thesis proposes ATFQ(Active Time Fairness Queuing)to guarantee the fairness among flows in multi-resource environments,as well as maximizing the resource utilizations and the scheduling efficiency.In each scheduling loop,ATFQ will determine the numbers of packets of different flows to be scheduled so as to guarantee the fairness among flows.ATFQ maximizes the resource utilizations and the scheduling efficiency through minimizing the total processing time of packets.With different numbers of scheduled packets in each scheduling loop,ATFQ emphasizes on the fairness and the scheduling efficiency by different degrees.Taking into account the deadline requirements of flows deriving from soft real-time applications,this thesis proposes MDFS(Multi-resource&Deadline-driven Flow Scheduling)to guarantee the deadlines of flows in multi-resource environments.While reducing the resource contention among flows,MDFS can also guarantee that the scheduled flows will never miss their deadlines.Secondly,in service chains data flows need to pass through multiple network function devices and undergo different kinds of function processing.The service difference between these devices makes it hard to guarantee the flow transmission on its routing path.This thesis proposes MRouting(Multi-resource Routing)to guarantee the deadline requirements of flows in service chains.Bear in mind that multiple paths exist between the flow sender and receiver,MRouting uses an appropriate routing selection scheme to select paths for newly arrived flows with the purpose of making new flows and previous flows finish their transmissions before their deadlines,as well as improving the network throughput.Finally,some cluster applications in data centers follow the coflow transmission pattern.A coflow is defined as a set of related parallel flows between two groups of servers.When these flows pass through different network function devices,the transmission requirements of coflows are difficult to be guaranteed.This thesis proposes DRGC(Data Rate Guarantee for Coflow)to guarantee the transmission rates of coflows in this situation.The priority tag is used to identify different coflows at the packet level,as well as determining their scheduling sequence.The timestamps of packets ensure that all the coflows only use necessary resources to maintain stable transmission rates,but will not occupy needless resources.In summary,this thesis proposes corresponding scheduling schemes,at the levels of packets,flows,and coflows,to protect the flow transmission in multi-resource environments.As verified by theoretical analysis and experiments,these schemes can efficiently guarantee the fairness,the deadlines and the data rate requirements of flows in multi-resource environments.