Many services on the Internet are provided by multiple identical servers in order to improve performance and robustness. The number, the location and the distribution of servers affect the performance and reliability of a service. The server placement is, however, often determined based on the empirical knowledge of the administrators. This paper investigates issues of the server placement in terms of the service performance and the server load. We identify that a server selection mechanism plays an important role in server placement, and thus, evaluate different server selection algorithms. The result shows that it is essential to the robustness of a service to employ a mechanism which distributes service requests to the servers according to the measured response time of each server. As a case study, we evaluate the server selection mechanisms employed by different DNS (Domain Name System) implementations. Then, we show the effects of the different server selection algorithms using root-server measurements taken at different locations around the world.

In this paper, we discuss the validity of the multi-scale gamma model and characterize the differences in host-level application traffic with this model by using a real traffic trace collected on a 150-Mbps transpacific link. First, we investigate the dependency of the model (parameters α and β, and fitting accuracy ε) on time scale Δ, then find suitable time scales for the model. Second, we inspect the relations among α, β, and ε, in order to characterize the differences in the types of applications. The main findings of the paper are as follows. (1) Different types of applications show different dependencies of α, β, and ε on Δ, and display different suitable Δs for the model. The model is more accurate if the traffic consists of intermittently-sent packets than other. (2) More appropriate models are obtained with specific α and β values (e.g., 0.1 < α < 1, and β < 2 for Δ = 500 ms). Also, application-specific traffic presents specific ranges of α, β, and ε for each Δ, so that these characteristics can be used in application identification methods such as anomaly detection and other machine learning methods.

In this paper, we show the effectiveness of software shaping through evaluation of our extensions to the internet transport protocols, TCP (Transmission Control Protocol) and UDP (User Datagram Protocol). These extensions are aimed at efficient realization of bulk data transfer and continuous media communication. The extensions are to be used with resource reservation, a possible and promising approach to resolve transport issues that the current TCP/IP networks cannot support. Although it seems straightforward to utilize dedicated bandwidth set up via resource reservation, filling up the reserved pipe is not so trivial. Performance analysis shows that, by applying the traffic shaping extensions, not only is the reserved pipe easily filled up, but the timely data delivery required by continuous media communication is also provided. Our experiments with a real system also show that overheads introduced by the new extensions are small enough to permit their practical use. The extensions are implemented in the UNIX system kernel.