Rational Server Selection for Mobile Agents

Mobile agents have the ability to migrate through heterogeneous networks and execute at remote hosts. This ability can be exploited to improve the performance of agent based applications by transferring agents to less heavily loaded hosts or by adapting the client/server interactions to the network load. That is, rather than transferring raw data through a bottleneck link, a small sized agent migrates to the information source, filters the data, and returns the processed data. Not only agent based applications benefit from such optimisation strategies. As the strategies balance the load, they also increase the overall system throughput. Yet to efficiently implement them, agents must be able to monitor the performance of network and host resources. Moreover, agent-mobility comes at a cost, and therefore it is not clear from the outset whether or not there is any benefit in an agent transfer at all. Although it is quite apparent that mobile agents may benefit from performance information, approaches for providing such information have not attracted much attention in the literature. A similar problem which has recently attracted some attention is server selection in the Internet: the increasing importance of bulk documents, audio, and video file transfer have led to the implementation of mirror sites. When downloading large files users have the choice between different mirror sites distributed across the Internet. In general, the user’s mirror site selection is either based on geographic information (that s, he selects a site close to his access network), or it is based on experience. However, involving the user in the selection process is neither comfortable for the user nor is the geographic location of a site a suitable performance metric. Recently, this has motivated automatic server selection approaches which take performance metrics like link bandwidth and latency into account, rather than the geographic location. Server selection approaches predict the performance of network links based on measurements. Although the problem of mirror site selection is related to the mobile agent’s migration problem, mirror site selection approaches do not meet all requirements of the agents. First of all, mirror selection focuses on network performance and does not take into account processing resources. Secondly, none of today’s approaches addresses the problem of rational server selection: agents are idle while the server selection system processes their requests, and therefore server selection comes at a cost. Hence, from an agent's perspective, server selection is only beneficial if the expected performance gain is greater than the selection costs. Thus, agents will only access a server selection system if the utility function U is positive: δ − − = min d d U average where daverage is the average service time, dmin is the service time of a server recommended by a selection system, and δ is the agent idle time while the selection system processes its request. In this thesis a server selection approach, the Performance Server, is developed which meets the agent’s requirements. The proposed approach is evaluated through benchmarks. Moreover, the problem of rational server selection is analysed. The Performance Server Figure 1 Performance Server A Performance Server consists of a Core Server and a set of Sensors and Navigators. Figure 1 shows the architecture of the Performance Server. Sensors keep track of resources like for instance CPUs, main memory and network links, and forward the compiled data to the dedicated Core Server. The Core Server implements the basic functionality of the Performance Server, and collects and analysis the sample data. Navigators use the sample data and statistics stored at the Core Servers to estimate the performance of remote sites, and select destination systems in behalf of the agents. To meet the different resource requirements and working plans of both mobile agents and services, the number of Navigators is not limited. Therefore, an agent must make sure it selects a Navigator which meets its particular evaluation criteria. The Performance Server has been evaluated through benchmarks. The results show that agents can significantly benefit from performance information. Yet, the gain heavily depends on the heterogeneity of the system, and on the agents’ resources requirements. Since a centralised approach is not scalable, the Performance Server approach has been extended to a distributed system. This system comprises numerous Performance Servers, each of which is responsible for a certain domain. The servers retrieve and exchange performance information from all domains. The server selection problem The main problem for the development of a decision algorithm for the server selection problem lies in the fact that the agents neither know the server resource capacities nor the waiting time δ. Yet, it is assumed that they have a basic knowledge of the system heterogeneity. Thus, let R be the resource capacity distribution describing the probability that a randomly selected server has a given resource capacity. According to the definition of R, its average value E[R] is the resource capacity an agent expects when it selects a server randomly, or if the number of alternative servers is 1. To estimate dmin, the random distribution Rn is required, which gives the maximum resource capacity out of n randomly selected servers. Obviously, Rn is an order statistic distribution. E[Rn] is the average maximum server capacity if a server is selected out of n others. Based on E[R] and E[Rn] the estimation of daverage and dmin is straightforward. As for the resource capacities, the waiting time is unknown up front. Here, the waiting time is assumed to be distributed according to a random distribution F, i.e. δ is given by F. An agent must now determine a time span T after which it stops waiting for the selection system’s response. Thus, the utility function of server-selection is given as follows: