Effectiveness of Server Load Estimation by using Request File Size for Web Server Clusters

There have been proposed several dynamic load-balancing methods for web server clusters (1)–(3) . Under those methods, arriving requests are dispatched to an appropriate server (may be the least loaded server) by a dedicated equipment called a load-balancer as shown in Fig. 1. In order to realize an effective load-balancing, those methods have required an almost real-time server load information which could be obtained from an observation of client-server interactions or a frequent polling from the load-balancer to each server. However the observation provides an indirect server load information and the polling interval produces a time-lag. Such uncertain load information would cause a performance degradation of the web server cluster. To overcome that problem, this paper proposes a load-balancing method based on a server load estimation. Under the proposed method, the load-balancer selects an estimated least-loaded server, and then assigns the arriving request to that server. The estimation is based on a fact that the file transfer delay is proportional to the file size . The fact makes it possible to estimate each server load from the last request assigned time and the file size. In order to validate the effectiveness of the proposed method, we examine the mean response time of several methods as shown in Fig. 2. For comparison, we employ Pick-2 (2) and Least-Connection methods (3) as representatives of the dynamic methods, and Round-Robin method for the static one. Fig. 2 shows the mean response time of each method under the case where the number of servers is 30. Round-Robin shows the worst performance. The method selects a server according to a predetermined order without regard to the server load. When the selected server is unfortunately heavy loaded, the response time would increase. Previously proposed dynamic methods shows smaller response time than the static method. Under Pick-2, the server load information is renewed in a certain period. So that the information may be stale. Least-Connection treats just the number of connections. Some of them may be idle. The proposed method shows the best performance. Since our method estimates each server’s load and then selects the least-loaded server, it provides an ideal delay characteristic. Fig. 1. Request dispatching by a loadbalancer