Performance test and analysis for an adaptive load balancing mechanism on distributed server cluster systems

Abstract

As the next generation of Internet services requires more highly scalable and available server systems, the cost-effective cluster of a large number of distributed computers becomes a popular solution. In this paper, we investigate to design and develop a server load balancing mechanism on cluster architecture, called the ALBM cluster. In order to construct the more scalable and reliable Internet service system, the ALBM cluster system consists of independent but co-operable components. The ALBM cluster system supports adaptive load balancing among servers via its adaptive load balancer (ALB) component in Layer 4 level and Layer 7 level. The Management Station (M-Station) and the Administrator Console are in charge of cluster management, system configuration management, and performance counter management. The Node Service is a system-level agent that is deployed into a server node. The node management and cluster management are the major tasks of the agent. Beside, the ALBM cluster is a flexible open system whose features of functionality can change or be easily expanded without affecting the rest of the system. In this paper, we also present a set of our experimental results to compare the performance of the ALBM cluster with that of LVS scheduling cluster system. We compare performance results of the ALBM with RR, LC, and WLC scheduling algorithms of a LVS cluster in both of homogeneous and heterogeneous system environments.

Introduction

As the commercial high-speed network technologies become mature, many end users easily access WAN (Wide Area Network) through telephone or network lines. This trend makes high-speed Internet services have become very popular to the public in the world. A lot of new and innovative commercial services via high-speed Internet are introduced day by day in various areas, such as home shopping malls, on-line cyber education, virtual community, virtual movie theater, etc. Due to the worldwide features of Internet, a new and nameless site comes to fame and grows to the giant one, achieving a couple of million clients. To serve a huge amount of client requests with the better performance, the next-generation Internet service over high-speed Internet might require more highly scalable and available systems.

A cost-effective cluster of a large number of distributed computers [1], [2] based on Layer 4 (L4) network load balancers becomes a typical solution to the scalable and reliable Internet services [17]. The L4 network load balancers act as traffic managers that would direct IP traffics to the most appropriate healthy server in a cluster, ensuring a maximum performance and providing a reliable connection. Originally intended to distribute traffics to stateless Web servers in a cluster, this kind of traffic manager is able to balance traffic loads to firewalls, ISP links, and VPN gateways. The role of the traffic manager in the data center is critical to manage highly available applications. Furthermore, the mission-critical enterprise applications based on a Web-based architecture demand the issue of scalable network load balancing.

In order to serve more general and various next-generation Internet services (such as real-time stream communication, on-line instant messenger, network games, as well as Web-based information services), we developed a highly scalable and reliable cluster system, called the ALBM (adaptive load balancing mechanism) cluster system. The ALBM cluster system has an integrated architecture with the following three major features. First, as does a general management system such as NMS (network management system) [6], [7], it has a centralized Management Station (M-Station) and a number of management agents distributed over managed nodes. Second, as do software-based cluster systems, the cluster has node services that take care of membership management and other cluster-related tasks. The last feature, as does a cluster with network load balancers, is that it has a network load balancer. The ALBM cluster system may have a number of network load balancers that can be also automatically constructed as a cluster, called the adaptive load balancer (ALB) cluster. The ALBM cluster architecture is scalable and fault-tolerant; the more resources are needed, the more computers can be added dynamically without reconstruction of the system and any loss of performance Even if a node is out of order, the cluster system is able to continue to provide the services as it is always available.

In this paper, we present the architecture of the ALBM cluster system that provides adaptive server load balancing as well as cluster and traffic management. The ALBM cluster system is composed of application servers, ALB, and the M-Station. The ALB provides network-level server load balancing that changes the scheduling adaptively according to the node state information by means of node agents. In order to analyze the dynamics of server performance due to the workload, we perform several experiments with performance counters. After evaluating the target performance counters with experiments and the appropriate threshold values, we compare the performance results of the ALBM cluster with those of the general LVS (Linux Virtual Server) clusters using Round-Robin, Least Connection, and Weighted Least Connection scheduling algorithms.

The organization of this paper is as follows. The next section presents the architecture of the ALBM cluster system. In the following section, performance counters related to the server load balancing are introduced with some experimental results. Section 4 presents experimental results that compare the performance of the ALBM cluster system to that of the LVS cluster system. In the last section, we conclude this paper with a summary.