Elsevier

Computer Networks

Volume 34, Issue 3, September 2000, Pages 355-361
Computer Networks

Placement of network servers in a wide-area network

https://doi.org/10.1016/S1389-1286(00)00115-8Get rights and content

Abstract

Various computing servers are springing up all over the Internet. These servers are used by client computers for different information retrieval purposes. The location of such servers has a direct impact on the networking cost associated with requests and replies to and from the servers (e.g., bandwidth and processing power costs). Badly placed servers result in an unnecessary cost that can be avoided by a better placement strategy. This paper introduces the network server problem and presents an algorithm that can be used for selecting least-cost assignment of the servers, based on the number of hosts in the network, their load, and the number of hops data packets traverse from host to server. We develop a mathematical model for this problem and suggest a solution procedure for the developed model. Computational experiments show that the proposed algorithm offers an attractive performance.

Introduction

The growth of computer networks in the last few years, as manifested in the ubiquitous Internet, coupled with the overwhelming proliferation of the client–server model, has spurred a vast number of computer servers (e.g., file servers, database servers, application servers, etc.) all over the globe. The typical scenario is that a client computer on one side of the network contacts a server computer on another side. The request packet goes through one or more network nodes (e.g., routers and switches) until it reaches the server. The server answers the client’s request and sends the required information back to the client.

The placement of such network servers is critical as it has a direct impact on the networking cost (i.e., the required bandwidth and processing power for each request). A badly placed set of severs may result in an excessive amount of bandwidth wasted by request and response packets that have to be routed over more links than necessary. Given a set of servers (SERVERS) to place in a wide-area network with a given network topology, there clearly is an optimal placement of the servers that minimizes the bandwidth cost, as measured by the number of links that packets must traverse.

A typical application of this placement strategy is the placement of mirror sites for a database or data repository, as is typically done in the Internet. A mirror site is used to decrease the load on the main site. It is usually placed in an area where it is expected that many users are located. Another application is the use of file servers by diskless workstations, or for network file servers’ (NFS) or Andrew File System (AFS) purposes, where workstations get their files from a remote server. World-Wide Web (WWW) servers are another, even more common, example. A more recent application involves caching and proxy servers [11]. In this development, server devices are placed at various locations around the Internet for the purpose of decreasing the overall load on the Internet backbones.

The idea for this research arose in the context of supporting connectionless traffic in Asynchronous Transfer Mode (ATM), a connection-oriented network, see [8]. In such an environment, a public ATM network supports connectionless traffic (e.g., Internet Protocol (IP) packets) by providing an overlay of connectionless servers (CLSs) on top of an ATM network. The CLSs serve to shield IP applications from the ATM network by acting as conventional routers, directing packets from one CLS to another from source to destination. By carefully placing these CLSs around the network, the overall cost associated with packets hopping from one CLS to another can be minimized. For an implementation and a case study of connectionless services in a local ATM network, see the work by Burak [4].

The need for a careful placement of servers in a network was also inspired by the various research efforts in the design of CLSs for ATM, see for example [9], [10], [13]. To the author’s knowledge, only Stavrov et al. [14] have tackled the placement problem in which an application program, called Bnet, is used to study various server placement alternatives. However, the method used for finding the minimum-cost assignment is that of simulated annealing which is different from Tabu search, the method employed by this paper.

In this paper, we first state the problem and give its mathematical formulation, present an algorithm that solves the placement problem, and give some computational results, followed by conclusions.

Section snippets

The problem statement

Given a set of SERVERS to place in a network of a given topology and a given number of hosts, find the least-cost solution to the problem of placing the SERVERS around the network in such a way that the total cost, as measured by the number of hops, is minimized (see Fig. 1). More specifically, let W be the number of switches (or routers) in a given network. Let hi be the number of hosts connecting to the network through switch number i, thus H={h1,h2,…,hW} is the set of hosts attached to each

The mathematical model

The integer programming model may then be expressed asMinimizeZ=∑j=1W∑i=1Whici,jxi,j+∑i=1Wc2xi,i(1−ui)subject to∑j=1Wxi,j=1fori=1,…,W,∑i=1Wxi,i=S.

The objective function in (1) represents the total cost of installing the SERVERS, where the first summation term represents the cost of SERVER placement, and the second term represents the cost of not utilizing the installed SERVERS. The objective of this function is to find the minimum total cost that places S SERVERS around the network, taking into

The solution procedure

In this section, we present how to solve the problem. Back to the mathematical model shown in the previous section, the solution to the model is an assignment of 1’s and 0’s to the square matrix X, which is W×W. As we showed in the previous section, the problem is decomposed into two subproblems. The solution to the first subproblem is the hardest. There are only two trivial cases, namely, S=0 and S=W. One can easily see that when W becomes large, the problem becomes harder. The solution to

Computational experiments

In this section, we test FHA with some randomly generated problems. First, we begin by an example.

Example 1

Suppose we have a network with three switches (W=3). It is required to install two servers (S=2) to solve the traffic problem in the network. Number of hosts connected to the network through switch 1, 2, and 3 are 2, 3, and 4, respectively. The matrix P, which gives the number of hops required for a host to reach another host, is shown in Table 1. The cost parameters c0,c1, and c2 are 5, 2, and

Conclusions

In this paper, we have developed a mathematical model for solving the problem of server placement in a computer network. The model takes into account both the cost of installing the servers and the cost of connecting every host to the servers. An algorithm has been proposed to solve the developed model, which has shown optimal solutions for the test cases employed by the conducted experiments. There are many applications of the mathematical model developed in this paper, including the recent

Acknowledgements

The authors would like to thank the anonymous referees for their careful reading of the first version of the paper and their wise suggestions, which have been incorporated herein, and enhanced the presentation of the paper.

Mohammad A. Al-Fawzan is an assistant professor in King Abdulaziz City for Science and Technology, Riyadh, Saudi Arabia. He received his B.Sc. and M.S. in industrial engineering from King Saud University, Riyadh, Saudi Arabia in 1988 and 1992, respectively. He received his Ph.D. in systems engineering from King Fahd University of Petroleum and Minerals, Dharan, Saudi Arabia in 1997. His research interests are in the areas of mathematical programming, production planning, quality control, and

References (16)

There are more references available in the full text version of this article.

Cited by (5)

  • Placement of network resources in communication networks

    2003, Computer Networks

  • A tabu search based algorithm for minimizing the number of tool switches on a flexible machine

    2003, Computers and Industrial Engineering
    Citation Excerpt :

    Since then, tabu search has been applied to a wide range of problem settings. Example of applications may be found in Al-Fawzan and Hoymani (2000). The tabu search starts at some initial point and then moves successively among neighboring points.

  • The cost-constrained web replica placement design in computer networks - An approximate approach

    2008, 2008 23rd International Symposium on Computer and Information Sciences, ISCIS 2008

  • Optimum server selection in content distribution networks

    2006, GLOBECOM - IEEE Global Telecommunications Conference

  • A Server Selection Method in Content Delivery Networks

    2003, IEICE Transactions on Communications

Mohammad A. Al-Fawzan is an assistant professor in King Abdulaziz City for Science and Technology, Riyadh, Saudi Arabia. He received his B.Sc. and M.S. in industrial engineering from King Saud University, Riyadh, Saudi Arabia in 1988 and 1992, respectively. He received his Ph.D. in systems engineering from King Fahd University of Petroleum and Minerals, Dharan, Saudi Arabia in 1997. His research interests are in the areas of mathematical programming, production planning, quality control, and maintenance. He has published in several journals including European Journal of Operational Research, International Journal of Production Research, International Journal of Production Economics, Production Planning and Control, Computers and Operations Research, and Annals of Operations Research. He has been an adjunct professor for King Saud University, industrial engineering department, since September 1998. He is a member in the council of the Saudi Industrial Engineering Chapter, Riyadh.

Fahad A. Hoymany holds a B.Sc. and two M.Sc.’s, all in computer science from the US. He also holds an MBA and a Ph.D. in computer science from the University of Pittsburgh, PA, 1997. In 1987–1988, he was technical chairman of Gulfnet, an academic computer network for the Arab Gulf countries, where he developed a number of network applications and was in charge of connecting Gulfnet to BITNET. At the end of 1998, he was appointed to head the Internet project for providing Saudi Arabia with Internet connectivity. His main research interest is in the area of high-speed networks, especially IP/ATM integration and ATM architectures. His Ph.D. Thesis centered around the design of an ATM switch that is capable of switching IP packets at full line speed.

View full text
Copyright © 2000 Elsevier Science B.V. All rights reserved.