Distribuição de serviços de comércio eletrônico

Abstract

The growth of E-commerce in the last years has considerably increased the demand on the sites that provide such services. As a consequence, servers are frequently overloaded, reducing the quality of service and raising the latency observed by the users, which, in a worst case scenario, may compromise the success of the virtual business due to the custumers insatisfaction. Traditional solutions, such as hardware upgrade, apply just to centralized servers, and therefore do not target an important part of the problem that is the delay inherent to the networkchannels. A common strategy to work around this problem is to distribute the service, using multiple servers spread across the network at nodes closer to the clients. The use of Proxy/Cache servers and Content Distribution Networks (CDNs) are two well-known examples of efficientdistributed Web services. However, in both cases, the services provided handle exclusively static information, i.e., information that is not frequently updated. For e-Commerce applications, this assumption is rarely valid, since data handled by the servers is essentialy dynamic for instance, a products availability in a virtual store or an items retail value in an eletronic auction.The implementation of a distributed service network that handles such limitations depends, among others, on the development of resource placement strategies that are flexible enough to adjust to the extremely dynamic workloads associated with such genre of applications. In this dissertation, we propose and evaluate two fundamental approaches to address this issue. The first is based on a linear optimization model that uses the network traffic and the number of hops between clients and distributed servers as the main metrics for the cost function. The second uses heuristics derived from the methods employed in CDNs to isolatedly solve the problems of server selection and resource distribution. The experiments performed show that, although the solutions provided are optimal, the linear model presents severe restrictions concerning the execution time, depending on the volume of input data. The approximate methods are not limited to these constraints and the resource distributions generated show a reasonable performance, with total costs ranging from 20% to 80% worse than the optimal solution. On the other hand, the heuristics have greater sensibility to workload variations, and are harder to calibrate due to the large number of parameters they,employ.