CS 791 -- Fall 1999

CS 791: Project Suggestions, Part One

September 19, 1999

Prof. John Byers

Overview

The projects listed below give an overview of some appropriate and hopefully interesting projects for CS 791. All sketch a current research avenue that is relevant to one or more of the Case Studies we are covering in the class, and are cross-listed according to those topics. In all cases, the description of the project is just a brief sketch, although in some cases, I have thought in detail about the research problem stated. You are by no means required to choose one of the projects listed here, but your project should fit in clearly with the theme(s) of the course. If your current research already fits in with the themes of the course, it is acceptable to continue working on such a project, provided that you set specific goals which can be completed by the end of the semester.

I prefer that students work individually on projects, but I will allow students to work in groups of two, in which case I will expect the scope of the project to be considerably larger, and the finished product to be more polished than if you work alone. Please get started early on your projects and come talk to me frequently to discuss what you are working on. Teaching this course is one of my main responsibilities this semester, so I will have plenty of time to discuss your projects with you.

All students are required to hand in a clearly written, one or two page project description in class on October 7. Please make sure that you have discussed your project with me prior to submitting this description so that we don't have multiple projects on identical topics.

"Improved Skyscraper Broadcasting" (Real-Time / Multicast)

In the Hua and Sheu paper [HS 97], the authors present a new scheme for transmitting video-on-demand in metropolitan-area networks. Their technique judiciously uses "channels" to transmit portions of a popular video, where each channel corresponds to a separate segment of the transmission. Users maintain the invariant that they play out the contents of segment i while subscribing to channel i + 1. The novelty in their work lies in the choices of the segment sizes so as to minimize the number of channels, while keeping both the latency perceived by clients and client buffer sizes to a minimum. Using fast forward error correcting techniques, the number of channels needed can be compressed even further while simultaneously making the system more resilient to packet loss without much additional overhead at the clients.

"Exploring Internet Power Laws" (Modeling)

A paper at this year's SIGCOMM [FFF99] observes that recent maps of the Internet topology exhibit power laws, including the observation that the distribution of nodes with a given outdegree follows a power law, i.e. the degrees follow a heavy-tailed distribution. In contrast, existing tools to construct "Internet-like topologies" all use much more regular distributions, thus it can be argued that the graphs they generate are not representative. Two challenges are to work on graph generation tools which actually exhibit the properties observed by [FFF99] and (perhaps more importantly) to provide explanations for why they arise.

"Maximizing Utility of Transport on Sensor Networks" (Wireless)

In future wireless sensor networks, nodes will be expected to self-organize into an ad hoc network, gather information about their surroundings and relay that information back to a destination in an adjoining wired network. Reliable delivery of this data is made difficult by virtue of the fact that some nodes may be faulty, and all are operating under (very) limited power constraints, which may cause nodes to become incapacitated. An interesting challenge is to develop distributed transport algorithms which maximize the expected amount of valuable information safely transmitted onto the wired network before the sensor network fails. (Sensor networks are a particularly rich source of interesting problems that can be explored in simulation).

Various Projects on Multimedia Streams (Real-time / Future Service Models / Multicast)

A multimedia session typically consists of a diverse ensemble of concurrent streams (audio, video, file transfer, shared whiteboard)
which have distinct application-level requirements, i.e. audio is inelastic to delay and jitter while the shared whiteboard may be able to tolerate fluctuations of this kind. In order to enable the network to provide reasonable service for such an application, the application must be able to specify its requirements.

From a future service models perspective, open questions about multimedia applications include how best to specify these application-level requirements, and how the network should provision for concurrent connections with different QoS requirements.
From a multicast perspective, a partial solution for transport of multimedia streams is the use of multiple multicast groups. But understanding how best to lay out the transmission over a small number of groups has not been studied extensively.
Techniques (not widely known in the networking community) are known for generating a priority encoding of a collection of data items. When this encoding is transmitted over a lossy network, the items can always be recovered in order of their priority, provided that a sufficiently large fraction of the encoding has been received by the client. The application of this idea to an ensemble of streams in a multimedia session should be direct.

"Measurement Collection in the Internet" (Modeling)

At the BU/NSF IMIC workshop on 8/30, Sugih Jamin gave a talk advocating placement of measurement collection nodes inside the Internet to gather statistics about the usage of the Internet at large. His group has so far been focusing on the theoretical and practical aspects of the question of where to place these measurement nodes and how to gather the measurements efficiently. Another interesting question is to consider how effective a small number of measurement nodes can possibly be? For instance, what fraction of the nodes and edges of a "representative topology" (see pitfalls above) can be observed by running experiments from a set of k measurement collection nodes. What value of k is at the "knee of the curve" with respect to capturing a representative slice of the entire topology?