Research On Energy-efficient Scheduling Of Applications In Virtual Network Functions On OpenStack

Telecom operators nowadays are trying to accelerate the deployment of new network services,and there is an increasing inability for those proprietary hardware-based appliances and services to meet their needs.Under such circumstances,Network Functions Virtualization(NFV)was proposed.By leveraging standard IT virtualization technology,NFV allows network functions of many kinds to run on industry-standard high volume servers,switches and storage as a software which located in data centers,network nodes and the end user premises.NFV can bring so many benefits that it has become the hot spot of telecom operators' and IT vendors' recent research.As a huge amount of server resources and power are wasted in current NFV data centers during the off-peak time,the research on NFV's energy-efficient scheduling would contribute to its promotion and popularization.To solve the problem,this thesis proposes an energy-efficient scheduling algorithm for NFV applications,and designs a scheduling system on OpenStack.In this thesis,we start by monitoring and analyzing the resource and energy consumption in data center.First,we design a method to collect the resource utilization of servers and virtual machines,then use these data to validate the energy models.The analysis of the server's energy model finds that the major part of server's energy consumption is consumed by the server when it's idle.So the fewer server use to consolidate the virtual machines,the less energy consumption is.Since a NFV service is compose by a set of virtual network function components that run on virtual machines,the key of NFV energy-efficient research is to consolidate the virtual machines that run the NFV applications onto minimal number of servers in the data center.On this basis,we design and implement an energy-efficient scheduling algorithm which can minimize the number of power-on servers,using probabilistic functions.The algorithm is intended to reduce the number of servers needed to run NFV applications.To achieve this,we introduce a probabilistic assignment function which can maximize the server's resource utilization during allocation.Meanwhile,in the migration procedure,the server will try to get rid of all the virtual machines when its resource utilization is low,so that the server can be switched off to save energy.Otherwise,if not all the virtual machines on the server can be migrated,then the virtual machines are not migrated at all to reduce the impact migration procedure has on the quality of service.Furthermore,for the overloaded case,the choice of migrating virtual machine is also optimized to only migrate one virtual machine.Finally,the algorithm is tested in both real and simulation systems,using a load model based on actual throughput of NFV applications.The result shows that in simulation system,the algorithm can save about 30% energy compared to VMware DRS and 10% to EcoCloud;in real tests,it can save about 20% energy compared to VMware DRS,which effectively reduce the power consumption in data center.