| Internet is now developing at the speed of more than the Law of Moore. Various applications and services are in great need of highly efficient and more reliable system to support. In network core and application of essential tasks (eg. WWW, FTP, VOD), the problem that too much load of a single server cannot guarantee the normal operation of Internet with appearance. Cluster system is one of those solutions that can support these applications to develop constantly and its ratio of performance to price is very high. If we apply cluster system to it, we can solve many problems such as the function of the server and it's competation. Load balance is set up on the base of existing Internet systems. It provides effective ways to broaden the server's bandwidth, increase Internet data handling capacity and improve its flexibility and functions. Load balance is a strategy. To different layers of Internet system, it has different specific solutions. Its focuses on those goals,such as: rational distribution of tasks, solutions of Internet jam, providing nearby service for the clients, improving visiting quality, quickening the servers'reaction speed and avoiding a single node invalidation in crucial places.There have already been some research fruits in the field of cluster system for network services. Early work relies on Round-Robin DNS to distribute incoming connections across a cluster of servers. However, this method cannot balance the load evenly. The MagicRouter of Berkeley acts as a switchboard that distributes requests for network service to the individual nodes in the cluster. Although it improves the quality of load-balance, this centralized approach is not inherently scalable because it does not take into account the fact that the MagicRouter becomes a bottleneck at high loads. Furthermore, In the eNetwork system of IBM which only alleviates the bottleneck problem by reducing the functions of the bottleneck point.In view of the above systems, this paper presents our research work on a cluster system for network services. It also proposes an optimized scheme and its implementation---Loop Reply. The theory and process of Loop Reply is as follows: in the cluster server the key process is called the process of Loop Reply. The client start a request, No.1 server accept the request and then it decides whether it can deal with the request or forward the request according to the load of server. Let's suppose No.1 server forward the request and No.2 server answer it. According to TCP/IP theory, when No.2 server answer the request, it must disguise as No.1 server so that the client will think it's legal. This theory demand fact that when No.2 server forward the data package, it must know the client's IP address and Server's IP address. Therefore, the common IP Package head can't satisfy its need. We can solve this problem with IP-IP encapsulation and MAC reform technology.Once being forward, No.1 Server must remember this address and forward all the later data package to No.2 server. In the process the client does not know who provide the service. Howerver, No.1 and No.2 server apply uncommon IP forward mode, it must modify the stack of TCP/IP in the inner core.Our system is a Linux-based PC cluster. Through the modification of the TCP/IP protocol stack, each node in the cluster can balance the network load automatically. Our scheme employs a distributed architecture and abandons the traditional centralized model, which makes our scheme especially suitable for PC cluster. In our design, each node has both abilities of service and redirection. Each node records the load values of the other nodes, and the local and redirected connections. According to this information, each node decides whether to redirect a package or not. In this way, the requests of network services are dispersed to different nodes. An internal communication protocol ensures the synchronization and integrity of the records in different nodes. This protocol also has the capability to detect node-fault. For the sake of high avai...
|
PDF Full Text Request