To illustrate the benefits of SatelliteLab as a platform for Internet measurement, we present a small-scale study of TCP throughput over cellular links. TCP flows are likely to experi- ence highly fluctuating throughput because cellular links suffer from interference, spotty coverage, and poor signal strength. Today, little is known about the characteristics of such links and how they affect the behavior of TCP flows.
Using SatelliteLab, we ran a series of TCP transfers between a mobile laptop and a well- provisioned server. The laptop was equipped with a UMTS modem and communication software that always chose the available network with the highest data rate (GPRS/EDGE or UMTS). Each transfer ran for 30 seconds and was spaced 4 minutes after the previous one. For each transfer we recorded the average throughput.
Figure 7.16 shows the cumulative distribution function of our TCP throughput measure- ments over UMTS/GPRS for both the upstream and the downstream direction. We find that downstream flows can have a throughput of anywhere between 10 Kbps to 320 Kbps (an order of magnitude difference!). In contrast, upstream flows have a bi-modal distribu- tion. Half of them have very low throughput of up to 10 Kbps, whereas the other half have high throughput of over 250 Kbps. In cellular networks, the available throughput depends on whether UMTS is available or whether GPRS/EDGE is used as a fall-back. Also, the signal quality can greatly affect throughput. This simple experiment illustrates how SatelliteLab can be used to conduct measurement studies of new network environments.
0% 20% 40% 60% 80% 100% 0 50 100 150 200 250 300 350 Percentage of flows Throughput (Kbps) Upstream Downstream
Figure 7.16: Cumulative distribution function of TCP throughput over UMTS. TCP over cellular links can experience very different throughput based on whether UMTS is available or whether it has to fall-back to GPRS/EDGE.
7.5 Applications
7.5.3 Summary
We have used SatelliteLab to evaluate two networked systems and to perform a measure- ment study. Our experiments illustrate three key benefits of the SatelliteLab testbed:
1. Evaluations on SatelliteLab and PlanetLab can yield substantially different results. The differences stem from the additional heterogeneity added by satellites. These results shed new insight into the observed performance of the evaluated systems. Even at small scale, SatelliteLab allows networking researchers to evaluate their prototypes over highly heterogeneous networks.
2. SatelliteLab can be used to debug the performance and behavior of deployed Internet systems. When a system behaves in a surprising manner on the Internet, researchers can use SatelliteLab to recreate the network conditions required to reproduce and understand performance issues of deployed systems.
3. SatelliteLab can be used as a measurement testbed for observing characteristics of different network environments, including wireless and broadband networks.
8 Conclusion and Future Work
In this section, we describe the high-level contributions of this thesis and discuss potential future research directions.
8.1 Summary
In the course of this thesis we developed a number of systems and tools designed to provide more transparency in broadband access networks for users, developers, and researchers.
We started by developing a novel measurement methodology that allows the study of broadband networks with minimal end host cooperation. In contrast to previous mea- surement tools, which have often required control over the measured host thus limiting the number of hosts that can be measured, our technique enabled us to study broadband access networks at scale for the first time. While we used this methodology to study broadband networks, the methodology can be used to study other types of networks as well.
We used this methodology to perform the first large-scale study of the characteristics of residential broadband access networks of major ISPs in Europe and North America. We characterized bandwidth, latencies, and loss rates of broadband networks and were able to show important differences between broadband and academic networks. For instance, we were the first to point out that many broadband hosts have deployed surprisingly long router queues, which can significantly affect the performance of latency-sensitive applications, such as VoIP and VoD.
Next, we developed the Glasnost system that allows users to test their access links for traffic differentiation. One of our design principles for Glasnost was to make it easy- to-use. We believe that this is one of the reasons that Glasnost was able to attract hundreds of thousands of users to date. The success of Glasnost inspired M-Lab, a platform to deploy measurement tools to enhance network transparency, which is supported by Google, PlanetLab, and other researchers. While our original version of Glasnost focused on the detection of blocking and throttling of BitTorrent traffic, we designed Glasnost to be extensible, allowing users to create and run their own Glasnost tests for arbitrary application traffic.
Using the data collected by Glasnost, we conducted the first large-scale study of the prevalence of traffic differentiation in broadband access networks. We were able to identify a number of major ISPs, including Comcast and Cox in the USA and StarHub in Singapore, that blocked BitTorrent traffic by injecting TCP RST packets into the transfer. Our data indicates that most ISPs stopped blocking BitTorrent traffic shortly after our results were widely covered in the media and caught the attention of telecommunication regulators. Since then, we found that ISPs rather than blocking increasingly throttle BitTorrent traffic. We then turned our attention to realistic evaluation of protocols and systems in broad- band networks. To evaluate and study transport protocol behavior at large scale over
diverse Internet paths, we developed Monarch. Monarch allows researchers to emulate transport protocol flows to a large number of hosts on the Internet without requiring di- rect access to them. This enables researchers to study the behavior of transport protocols in the wild, and in particular in broadband networks. For instance, using Monarch we were able to show that the large router queues that are common in broadband networks can lead to high loss rates and high packet latencies. This can be problematic for widely deployed TCP congestion control algorithms that interpret loss events as an indication of congestion.
Finally, we presented SatelliteLab, a new testbed design that makes it easy to add broadband nodes to existing Internet testbeds such as PlanetLab. SatelliteLab solves the problem that today’s testbeds mostly consist of well-connected academic nodes, and thus do not cover the heterogeneity of the Internet. In fact, given the high requirements for testbed nodes, it is often not possible for broadband nodes to join a testbed. SatelliteLab makes it possible to supplement testbeds with arbitrary nodes, including broadband nodes and even smartphones, thus vastly increasing their heterogeneity. Using SatelliteLab, we were able to identify several issues with an overlay multicast system and a network coordinate system in broadband environments that did not occur when evaluating these systems in state-of-the-art testbeds.