No abstract

This paper presents the concept of change detection forfilter-based network-state estimation. This could be useful in various contexts; two examples are network management and adaptive applications. In particular, it is shown that the performance of availablebandwidth estimation can be significantly enhanced by employing a change-detection technique in conjunction with a filter-based estimator. By using filter-based approaches, it is possible to track the state of communication systems, and to estimate network properties in real-time. A virtue of filter-based methods is the ability to enhance the estimation performance by combining them with change detection. This makes itfeasible to overcome the tradeoffs regarding speed of adaptation to changes versus stable estimation. We discuss filtering and change detection in general, and illustrate the power of this combination with the filter-based available-bandwidth estimator BART enhanced by the light-weight changedetection test CUSUM.

resolution may be different for various applications.Abstract-In this paper we present a way of tuning the The time scale of bandwidth tracking offered by BART is temporal characteristics of a new available-bandwidth related to two adjustable properties. The obvious one is the estimation method, BART. The estimation engine in this inter-probing time. In the example above, when we probe method is Kalman-filter based. A current estimate of the once per second, we cannot hope to accurately track the available bandwidth is maintained, and for each new sequence bandwidth fluctuations at any time resolution better than of probe packet pairs an updated estimate is produced. The main input parameters needed by the Kalman filter are the approximately two seconds (cf. the sampling theorem in variance of the measurement noise and the covariance of the information theory). By decreasing the inter-probing time, process noise. The former is measured by the method, whereas an improved tracking performance is expected; however, the the latter is not in general attainable by analytical or empirical trade-off is the larger amount of probe traffic affecting the investigation. Instead, it is reasonable to treat this as a tunable network. The more subtle adjustable property is the process parameter. We discuss how the temporal characteristics of the noise covariance, Q, one of the crucial filter parameters.tracking of end-to-end available bandwidth may be tuned. This is a 2x2 matrix in the BART formalism.In the present paper, we explore how the elements of Q could be chosen so as to optimize the tracking performance I. INTRODUCTION of BART for desired bandwidth variability. We introduce A. Overview and apply a specific variability measure, which captures the The capability of estimating end-to-end available effect of both time resolution and traffic aggregation. bandwidth is useful in several contexts, including service B. Related Work level agreement verification, network monitoring and server Several other bandwidth estimation methods have been selection. Estimation of bandwidth in real-time, with per-proposed [3, 4, 5, 6, 7, 8, 9, 10]. These methods have been sample update, opens up for many applications, including compared for performance [11, 12]. adaptation based on available bandwidth directly (rather In a seminal paper on packet-pair techniques [13], than measures such as loss or delay) in e.g. congestion Keshav discussed using a Kalman filter for the estimation of control and streaming of audio and video. "bottleneck service rate" for end-point flow control Estimation methods based on a filtering approach possess purposes, and concluded that this would not be practical. this compelling feature of producing an updated estimate for However, his analysis rested on the assumption that queuing each new sampling of the system properties. For instance, service in network nodes is based on stateful flow-based by probing the network with probe packets once every round-robin instead of stateless first-come first-served second, one ...