verprovisioning is widely used by packet network engineering teams to protect networks against network element failure and support the rapid growth of traffic volume. So far, this approach has been successful in maintaining simple, scalable, highly available, and robust networks. It is important to realize that in packet networks which do not perform call admission control, there is often no way to control the amount or types of traffic entering the network. The provisioning problem therefore lies in figuring out how much excess capacity is required to provide robustness (e.g., resilience to multiple simultaneous link failures) and scalability. The current tools for network management, such as Simple Network Management Protocol (SNMP), are limited in their capabilities, since they only provide highly aggregated statistics about the traffic (e.g., average traffic load over fiveminute intervals) and do not give insight into traffic dynamics on timescales appropriate for events such as packet drops. Another example is the demand traffic matrix, which is a crucial input to many network planning, provisioning, and engineering problems, but is difficult to obtain with available tools [1,2].Detailed traffic measurements are necessary to assess the capacity requirements and efficiently engineer the network. In this article we first describe the architecture and capabilities of the IPMON system. Then we point out the challenges we faced in collecting terabytes of data, and include our solutions to data sanitization. In the remainder of the article we AbstractNetwork traffic measurements provide essential data for networking research and network management. In this article we describe a passive monitoring system designed to capture GPS synchronized packet-level traffic measurements on OC-3, OC-12, and OC-48 links. Our system is deployed in four POPs in the Sprint IP backbone. Measurement data is stored on a 10 Tbyte storage area network and analyzed on a computing cluster. We present a set of results to both demonstrate the strength of the system and identify recent changes in Internet traffic characteristics. The results include traffic workload, analyses of TCP flow round-trip times, out-of-sequence packet rates, and packet delay. We also show that some links no longer carry Web traffic as their dominant component to the benefit of file sharing and media streaming. On most links we monitored, TCP flows exhibit low out-of-sequence packet rates, and backbone delays are dominated by the speed of light.

Abstract-To support latency sensitive traffic such as voice, network providers can either use service differentiation to prioritize such traffic or provision their network with enough bandwidth so that all traffic meets the most stringent delay requirements. In the context of widearea Internet backbones, two factors make overprovisioning an attractive approach. First, the high link speeds and large volumes of traffic make service differentiation complex and potentially costly to deploy. Second, given the degree of aggregation and resulting traffic characteristics, the amount of overprovisioning necessary may not be very large. This study develops a methodology to compute the amount of overprovisioning required to support a given delay requirement. We first develop a model for backbone traffic which is needed to compute the end-to-end delay through the network. The model is validated using 331 one-hour traffic measurements collected from the Sprint IP network. We then develop a procedure which uses this model to find the amount of bandwidth needed on each link in the network so that an end-to-end delay requirement is satisfied. Applying this procedure to the Sprint network, we find that satisfying end-to-end delay requirements as low as 3 ms requires only 15% extra bandwidth above the average data rate of the traffic.

Abstract-We measure and analyze the single-hop packet delay through operational routers in a backbone IP network. First we present our delay measurements through a single router. Then we identify stepby-step the factors contributing to single-hop delay. In addition to packet processing, transmission, and queueing delays, we identify the presence of very large delays due to non-work-conserving router behavior. We use a simple output queue model to separate those delay components. Our step-by-step methodology used to obtain the pure queueing delay is easily applicable to any single-hop delay measurements.After obtaining the queueing delay, we analyze the tail of its distribution, and find that it is long tailed and fits a Weibull distribution with the scale parameter, ¼ , and the shape parameter, ¼ to ¼ . The measured average queueing delay is larger than predicted by M/M/1, M/G/1, and FBM models when the link utilization is below 70%, but its absolute value is quite small.

Abstract-We measure and analyze the single-hop packet delay through operational routers in the Sprint Internet protocol (IP) backbone network. After presenting our delay measurements through a single router for OC-3 and OC-12 link speeds, we propose a methodology to identify the factors contributing to single-hop delay. In addition to packet processing, transmission, and queueing delay at the output link, we observe the presence of very large delays that cannot be explained within the context of a first-in first-out output queue model. We isolate and analyze these outliers.Results indicate that there is very little queueing taking place in Sprint's backbone. As link speeds increase, transmission delay decreases and the dominant part of single-hop delay is packet processing time. We show that if a packet is received and transmitted on the same linecard, it experiences less than 20 s of delay. If the packet is transmitted across the switch fabric, its delay doubles in magnitude. We observe that processing due to IP options results in single-hop delays in the order of milliseconds. Milliseconds of delay may also be experienced by packets that do not carry IP options. We attribute those delays to router idiosyncratic behavior that affects less than 1% of the packets. Finally, we show that the queueing delay distribution is long-tailed and can be approximated with a Weibull distribution with the scale parameter = 0 5 and the shape parameter = 0 6 to 0.82. Index Terms-Link utilization, queueing delay, single-hop delay measurement.

No abstract