Application-level routing that chooses an end-to-end traffic route that relays other end hosts can improve user-perceived performance metrics such as end-to-end latency and available bandwidth. However, selfish route selection performed by each end user can lead to a decrease in path performance due to overload by route overlaps, as well as an increase in the inter-ISP transit cost as a result of utilizing more transit links compared with native IP routing. In this paper, we first strictly define an optimization problem for selecting application-level traffic routes with the aim of maximizing end-to-end network performance under a transit cost constraint. We then propose an application-level traffic routing method based on distributed simulated annealing to obtain good solutions to the problem. We evaluate the performance of the proposed method by assuming that PlanetLab nodes utilize application-level traffic routing. We show that the proposed routing method can result in considerable improvement of network performance without increasing transit cost. In particular, when using end-to-end latency as a routing metric, the number of overloaded end-to-end paths can be reduced by about 65%, as compared with that when using non-coordinated methods. We also demonstrate that the proposed method can react to dynamic changes in traffic demand and select appropriate routes.

Overlay routing is an application-level routing mechanism on overlay networks. Previous researches have revealed that the overlay routing can improve user-perceived performance. However, it may also generate traffic unintended by ISPs, incurring additional monetary cost. In addition, since ISPs and end users have their own objectives respectively regarding traffic routing, overlay routing must be operated considering both standpoints. In the present paper, we propose a method to reduce inter-ISP transit costs caused by overlay routing from the both standpoints of ISPs and end users. To determine the relationships among ASes, which are required for ISP cost-aware routing, we construct a method to estimate a transit cost of overlay-routed paths from end-to-end network performance values. Utilizing the metric, we propose a novel method that controls overlay routing from the both standpoints of ISPs and end users. Through extensive evaluations using measurement results from the actual network environments, we confirm that the advantage of the proposed method whereby we can reduce the transit cost in the overlay routing and can control the overlay routing according to the objectives of both ISPs and end users.

Predicting network throughput is important for network-aware applications. Network throughput depends on a number of factors, and many throughput prediction methods have been proposed. However, many of these methods are suffering from the fact that a distribution of traffic fluctuation is unclear and the scale and the bandwidth of networks are rapidly increasing. Furthermore, virtual machines are used as platforms in many network research and services fields, and they can affect network measurement. A prediction method that uses pairs of differently sized connections has been proposed. This method, which we call connection pair, features a small probe transfer using the TCP that can be used to predict the throughput of a large data transfer. We focus on measurements, analyses, and modeling for precise prediction results. We first clarified that the actual throughput for the connection pair is non-linearly and monotonically changed with noise. Second, we built a previously proposed predictor using the same training data sets as for our proposed method, and it was unsuitable for considering the above characteristics. We propose a throughput prediction method based on the connection pair that uses ν-support vector regression and the polynomial kernel to deal with prediction models represented as a non-linear and continuous monotonic function. The prediction results of our method compared to those of the previous predictor are more accurate. Moreover, under an unstable network state, the drop in accuracy is also smaller than that of the previous predictor.

Network testbeds have been used for network measurement and experiments. In such testbeds, resources, such as CPU, memory, and I/O interfaces, are shared and virtualized to maximize node utility for many users. A few studies have investigated the impact of virtualization on precise network measurement and understood Internet traffic characteristics on virtualized testbeds. Although scheduling latency and heavy loads are reportedly affected in precise network measurement, no clear conditions or criteria have been established. Moreover, empirical-statistical criteria and methods that pick out anomalous cases for precise network experiments are required on userland because virtualization technology used in the provided testbeds is hardly replaceable. In this paper, we show that ‘oversize packet spacing’, which can be caused by CPU scheduling latency, is a major cause of throughput instability on a virtualized network testbed even when no significant changes occur in well-known network metrics. These are unusual anomalies on virtualized network environment. Empirical-statistical analysis results accord with results at previous work. If network throughput is decreased by the anomalies, we should carefully review measurement results. Our empirical approach enables anomalous cases to be identified. We present CPU availability as an important criterion for estimating the anomalies.