Ethan Blanton

On the impact of bursting on TCP performance

Periodically in the transport protocol research community, the idea of introducing a burst mitigation strategy is voiced. In this paper we assess the prevalence and implications of bursts in the context of real TCP traffic in order to better inform a decision on whether TCPs congestion control algorithms need to incorporate some form of burst suppression. After analyzing traffic from three networks, we find that bursts are fairly rare and only large bursts (of hundreds of segments) cause loss in practice.

Notes on burst mitigation for transport protocols

In this note we explore the various causes of micro-bursting in TCP connections and also the behavior of several mitigations that have been suggested in the literature along with extensions we develop herein. This note methodically sketches the behavior of the mitigations and presents the tradeoffs of various schemes as a data point in the ongoing discussion about preventing bursting in transport protocols.

Resource management in an active measurement service

Many network services such as voice, video, and collaborative applications require an informed view of network characteristics for effective operation. Sharing a network measurement service across multiple applications can significantly reduce measurement overhead, and remove the burden of performing network measurements from individual applications. To be effectively shared, a network measurement service must provide a variety of measurements to applications on-demand, including end-to-end available bandwidth, delay, and loss. We propose such a service, with a focus on quantifying and bounding the impact of active measurements on the network resources being measured. Resource bounds are necessary for wide-scale deployment, since not all users of the service can be trusted. The service informs applications of how service bounds affect their measurements. We introduce methods to characterize the behavior of active measurements for use in admission control and scheduling decisions. We evaluate our methods in experiments under realistic scenarios on the Emulab testbed.

Design and evaluation of the S 3 monitor network measurement service on GENI

Network monitoring capabilities are critical for both network operators and networked applications. In the context of an experimental test facility, network measurement is important for researchers experimenting with new network architectures and applications, as well as operators of the test facility itself. The Global Environment for Network Innovations (GENI) is a sophisticated test facility comprised of multiple “control frameworks.” In this paper, we describe the design and implementation of S3 Monitor, a scalable and extensible monitoring service for GENI. A key feature of S3 Monitor is a flexible design that allows easy “plug in” of new network measurement tools. We discuss our deployment experiences with S3 Monitor on GENI, and give experimental results to quantify the performance and system footprint of S3 Monitor. We find that the S3 Monitor service is light-weight and scales well as the number of paths to be monitored increases.

On the Utility of Inference Mechanisms

A number of network path delay, loss, or bandwidth inference mechanisms have been proposed over the past decade. Concurrently, several network measurement services have been deployed over the Internet and intranets. We consider inference mechanisms that use O(n) end-to-end measurements to predict the O(n^2) end-to-end pairwise measurements among n nodes, and investigate when it is beneficial to use them in measurement services. In particular, we address the following questions: (1) For which measurement request patterns would using an inference mechanism be advantageous? (2) How does a measurement service determine the set of hosts that should utilize inference mechanisms, as opposed to those that are better served using direct end-to-end measurements? (3) How can the answer to question 2 be efficiently computed as measurement requests arrive and terminate? Our solution is able to identify groups of hosts which are likely to benefit from inference, by utilizing a probabilistically generated spanning forest on the measurement request graph. We compare our solution to a simple heuristic that uses the number of measurements a host participates in. Results with synthetic datasets as well as datasets from a popular peer-to-peer system demonstrate that our technique identifies host subsets that benefit from inference quite accurately, and in significantly less time than an algorithm that identifies optimal subsets. The measurement savings are large when measurement request patterns exhibit small-world characteristics, which is often the case for peer-to-peer and other popular distributed systems.

On the Cost of Network Inference Mechanisms

A number of network path delay, loss, or bandwidth inference mechanisms have been proposed over the past decade. Concurrently, several network measurement services have been deployed over the Internet and intranets. We consider inference mechanisms that use O(n) end-to-end measurements to predict the O(n2) end-to-end pairwise measurements among n nodes, and investigate when it is beneficial to use them in measurement services. In particular, we address the following questions : 1) For which measurement request patterns would using an inference mechanism be advantageous? 2) How does a measurement service determine the set of hosts that should utilize inference mechanisms, as opposed to those that are better served using direct end-to-end measurements? We explore three solutions that identify groups of hosts which are likely to benefit from inference. We compare these solutions in terms of effectiveness and algorithmic complexity. Results with synthetic data sets and data sets from a popular peer-to-peer system demonstrate that our techniques accurately identify host subsets that benefit from inference, in significantly less time than an algorithm that identifies optimal subsets. The measurement savings are large when measurement request patterns exhibit small-world characteristics, which is often the case. (Part of this work (focusing on one of three solutions presented in this paper) appeared in).

Location matters: Eliciting responses to direct probes

In this work, we propose techniques to attain visibility into an arbitrary Internet subnetwork that is responsive to indirect probes but not to direct probes. By probing the network from a small number of selected vantage points, we are able to collect information about network-layer topology which would otherwise be hidden from measurement due to rate limiting practices, security mechanisms, and routing dynamics. We investigate the reasons for differing visibility, and the required number and placement strategies of vantage points needed to collect topology information at a low cost. We demonstrate substantial improvement in global visibility as probed by the TraceNET path measurement tool when leveraging only five vantage points selected according to route similarity.

Controlling the cost and increasing the utility of network measurement infrastructures

Network measurement is an important tool in network research and operations, and is increasingly utilized in end-user applications. However, active measurements compete with other applications for limited network resources. These measurements are typically performed by individual users and applications, with limited coordination or planning. This dissertation addresses network measurement in the context of a measurement infrastructure, which can coordinate measurements, monitor and control their resource consumption, and even improve the quality of measurement results. We address three specific areas of operation in a measurement infrastructure. We present an architecture for an infrastructure that uses admission control to bound the resource consumption of measurements. We use estimates of the resources consumed by real-world measurement tools in the admission control mechanism. We evaluate methods of estimating the resource consumption of measurements in terms of network bandwidth, and show that these estimates are important to the accuracy of admission control. We propose several algorithms for determining when network measurements can be replaced by a network inference mechanism to reduce measurement load in the network at a cost of measurement accuracy. We experimentally evaluate these algorithms on measurement patterns designed to present difficulties, measurement patterns based on real-world peer-to-peer distributed communication, and random measurements. We show that the algorithms successfully reduce the measurement load on the first two types of graphs, and that they do not incur undue expense on the latter. We design and evaluate a family of methods for selecting vantage points from which to take active measurements to increase reachability into the Internet. We show that two of these methods perform substantially better than selecting vantage points at random, allowing a larger number of hosts to be probed with direct measurements with the same number of vantage points.