The Research And Implementation Of Cluster Network Services Based On Proactive Schedule

The dramatic evolution of Internet has changed human's life deeply in recent twenty years. Network service is the main form of Internet applications. Facing rapidly increasing number of user, more complicated application and maturation of service management, it must be up against unprecedented pressure and challenges.People have done a lot and achieved to construct high-performance, highly scalable and highly available network services with cluster technologies. However, with the increasing deployment of key applications like e-government and e-commerce on Internet, more complex and challenging problems arose. They include QoS control, full utilization of system resource, and cost-effectiveness of network service. QoS control plays a key role among the above problems but has not been well resolved. The work addressed in this thesis represents a step towards ameliorating this situation.Based on the analysis of the requirements for network application and current cluster architectures for network services, we found that the cluster architecture based on load balancing can't come up with the new requirements of network services such as quantitative QoS control. New cluster architecture focusing on QoS control should be brought forward. Modern control theory and technique have provided important guidance and tools for this job.According to the basic architecture of a feedback control system, we present a new cluster architecture for network services based on proactive schedule mechanism in which each component corresponds to a counterpart in the feedback control system and the request schedule process corresponds to the control process. It's featured with decentralized control approach against the extensibility problem in a central control structure of a traditional cluster. The QoS controllers of the network service run on each server node in the cluster to ensure enough resource for the requirements of QoS control on performance, availability and feasibility. We call this a proactive scheme because each server makes decisions for itself. Analysis based on system model showed that the decentralized control approach has better performance than a traditional central control one.QoS control usually needs a 7-layer schedule based on request classification which requires more system resources. To improve the extensibility and availability of the cluster system, we propose distributed cluster architecture with multiple entries and corresponding fully distributed proactive schedule mechanism. Then we give two applicable solutions for constructing large scale proactive cluster system without extensibility and availability problems.The design of quantitative QoS control subsystem is the key problem for proactive cluster system. Due to the extraordinary complexity of network applications, quantitative QoS control, especially for a large scale heterogeneous cluster, is a big challenge. An applicable solution should be not only effective and adaptable but also compatible with existing system and easy to deploy and use. Traditional design methods such as those based on experience, analytical model or PID control are not competent for this job. We have tried to apply modern intelligent control methods to the design of quantitative QoS controller.Fuzzy control is the most well-developed control method among modern intelligent control methods. Based on fuzzy control theory, we have designed and implemented four fuzzy controllers with four different typical control objectives to control the average response time, data throughput, response time ratio, data throughput ratio of different request types. Experiments verified the effectiveness of these controllers on response speed, convergence and stability.Considering fuzzy controller still has some disadvantages on steady state error and adaptability, we bring forward an adaptable control method with the help of artificial neural network (ANN). Combining ANN and traditional PID control, we propose an adaptable PIDcontroller based on BP network. BP network can help the PID controller to track the changing environment and system under control using its online learning capability. Simulation proved the good effectiveness of BP-PID controller. We also designed and implemented four BP-PID controllers for four different typical control objectives as mentioned above. Tests verified their steady state performance and adaptability in contrast to their fuzzy counterparts.The response stage of a BP-PID controller is not as steady as a fuzzy controller. So we try to combine the strong points of ANN and fuzzy system to design an ANN-Fuzzy controller, using ANN to adapt the fuzzy system according to the change of the environment and the system under control. Simulation proved the advantage of ANN-Fuzzy system over a simple BP network. We also designed and implemented four ANN-Fuzzy controllers for four different typical control objectives as mentioned above. Test results showed that ANN-Fuzzy controllers exceed their counterparts designed with fuzzy controller or BP-PID controller on both steady state performance and dynamic state performance.Finally, based on the above work, we designed and implemented an applicable proactive cluster system, which includes:Distributed framework featured with decentralized quantitative QoS control mechanism and multiple cluster system entries. It has very high extensibility and availability.High-performance, highly scalable application gateway designed and implemented based on Linux kernel. It supports many application protocols such as HTTP, E-mail. It also supports the development and deployment of other new application protocols.Four basic controller types: basic PID controller and its variations, fuzzy controller, BP-PID controller, ANN-Fuzzy controller.16 typical quantitative QoS controllers designed with the above basic controller types. They are used to control the average response time, data throughput, response time ratio, data throughput ratio of different request types.Highly extensible quantitative QoS control framework. It supports the development and deployment of other new controller types. It also support many OSes (Windows series and Unix-like series).Related cluster management software, providing versatile managing interfaces including Web, terminal and command line.Various testing have verified the effectiveness and practicability of the proactive cluster system. The product using proactive cluster system has also been brought forward to the market and deployed in many places, such as KingSoft's websites.