AT&T Labs Research
(http://www.research.att.com)
Welcomes
AT&T Labs 2006
University Collaborations Symposium
http://www.research.att.com/ucollaborations
Shannon Laboratory 180 Park Avenue - Bldg.
103
Florham Park, NJ
August 3 & 4, 2006
Florham Park Auditorium, C050
Middletown - Conference Room A5-5A02 (with video link)
Austin, Texas - Room 220 at 9505 Arboretum Boulevard (with video link)
San Ramon, California - Toyon Room at 2600 Camino Ramon (with video link)
Contact
Information
Mahmoud
Daneshmand
AT&T Labs Research
Director, University Collaborations
daneshmand@research.att.com
MIDDLETOWN 732-420-4124
FLORHAM
PARK 973-360-8604
ABSTRACTS
1. LiveRAC: Live
Reorderable Accordion Drawing
2. Dragnet: Towards
a Framework of Internet Forensics
3. Modulation
Envelopes of Speech Signals Can't Be Strictly Positive
4. Time-Frequency
Reassignment and Modulation
5. Verbal Query for IP-TV: A Natural Interface to the World of
Entertainment
6. Research and
Benchmarking of Advanced Video Search Technologies
7. Statistical
Techniques for Network Data
8. A P2P
Distributed Storage Service for the AT&T Lightspeed Network
9. Enabling
AT&T 4G Systems
10. Multi-User
and Multi-Cellular MIMO Wireless Communication
11. Failure
Management in Commercial Data Centers
12. Towards a
Free-Viewpoint Video System
13. Next Generation
Video Coding and Image/Video Quality Assessment Algorithms
14. Incorporating
Prosody in Speech Translation
15. Using Tree-Drawing
Algorithms to Visualize Tree-Like Networks
16. Xtreenet: A
Scalable Unified Overlay Network for XML Content Access and
Distribution
17. Distributed
Route Control for Customer Traffic Load-Balancing
18. Emulating
Large Scale Enterprise Networks
19. Simulation
and Analysis of Failures and Impairments for Multimedia Distribution
Networks
20. Toward a Fair
and Robust Internet Backbone Routing
21. Multi-Level
Raid for Very Large Disk Arrays - VLDAs
22. Troubleshooting
in MPLS Virtual Private Networks
23. Technology
Diffusion and Long Term Forecasting: Application to Growth of Wireless Mobile
Services
24. Robust and
Loss-Tolerant Link and Transport Protocols for Wireless Network
Environments
25. Underwater
Sensor Networks: Applications and Challenges
26. Experiments
and Performance of RFID Systems
27. Reliability
in RFID Systems
28. Mobility of
Inventors and Growth of Technology Clusters: An Examination of Innovation
Networks in the Telecom Industry
29. Optimizing
Large File Transfers with WORMHOLE
30. Better Traffic
Matrix Estimation in a Large ISP Network
31. Toward Fully
Automated Verification and Optimization of the RUBY System
32. Use of Communication Graphs for Real-Time Anomaly
Detection
33. Anticipating
Customer Behavior through Traffic Analysis
34. Darkstar - Anomaly Detection
ABSTRACTS
1. LiveRAC: Live Reorderable Accordion Drawing
Peter McLachlan, University of British
Columbia
Tamara Munzner, University of British Columbia
Eleftherios Koutsofios, AT&T Labs Inc. - Research
Stephen North, AT&T Labs Inc. - Research
Abstract:
LiveRAC is a focus+context approach for monitoring systems and networking data which addresses the challenges of providing scalability, user directed data reordering, and details-on-demand. LiveRAC maps alarm and metric data from network devices including servers, routers and switches into a visual metaphor called accordion drawing. In accordion drawing, users interact with the display as though it were a rubber sheet tacked down at the borders. Regions of the display can be stretched and compressed, but the fixed borders ensure the visibility of the entire information space. Compressed regions of the display aggregate the underlying data. We implement visual landmarks, a mechanism for ensuring the visibility of important features such as critical alarms. LiveRAC extends existing accordion drawing techniques in two ways: it can add and remove objects to the data set in logarithmic time, and provides a framework for semantic zooming, a visualization technique where an optimal data representation is selected based on the screen space available to a data cell. Using a client-server approach we allow the user to query the underlying data in a context rich, visually salient metaphor while maintaining interactive frame rates.
2. Dragnet: Towards a Framework of Internet Forensics
Hui Zhang, School of Computer Science, Carnegie Mellon University
Abstract:
The world of network security is an arms race where attackers constantly change the signatures of their attacks to avoid detection. In addition, all sophisticated attackers use multi-level attacks that begin by compromising or infecting innocent hosts and then use these zombies to initiate and propagate the actual attack. In the Dragnet project, we study forensic techniques that can identify the true sources of attacks and reconstruct the initial attack propagation events. We believe this capability can significantly strengthen the hand of administrators in deterring attacks or correcting the weak points in a network perimeter.
I will first describe a novel ``random moonwalk'' algorithm that automatically pinpoints the origin of an epidemic spreading attack such as Internet worms and its initial successful infection events. The algorithm is agnostic to attack specific characteristics such as payload contents, port numbers used, or specific software vulnerabilities exploited. It is effective in identifying the origins of both today's fast propagating worms and a wide class of stealthy worms that attempt to hide their attack flows among background traffic. I will further describe Dragnet in federated network environments such as the Internet, where multiple administrative domains jointly perform epidemic forensic analysis without releasing information that is not available to each other. The federated Dragnet system not only achieves comparable performance to centralized forensics, but also is incentive compatible: each domain's own forensics capabilities are enhanced by participating, even in partial deployment scenarios.
3. Modulation Envelopes of Speech Signals Can't be Strictly Positive
Les Atlas and Bishnu Atal, Department of Electrical Engineering, University of Washington
Abstract:
There is substantial evidence that many natural signals, such as speech, can be represented as low frequency modulators which modulate higher frequency carriers. For example, the low-frequency modulators are considered to represent a key dimension used by humans when they separate multiple simultaneous conversations. For decades, researchers have attempted to get at this modulation representation by breaking signals into frequency sub-bands, finding positive and real envelopes for each sub-band, and filtering or transforming these temporal envelopes. Unfortunately, these previous modulation filtering operations have little quantitative justification. A correct and substantive definition of modulation filtering would potentially advance signal processing, enhancement, and representation for many applications, especially speech processing.
With few, if any, exceptions, all past researchers have assumed that the sub-band envelopes are non-negative and real. These past assumptions allow the modulator/carrier separation to be done incoherently, as in a simple AM radio. Spectral symmetry arguments can show that these incoherent, positive and real modulator assumptions are rarely satisfied for real speech and audio signals. By forcing this incorrect assumption, almost any modification of these incorrectly formed envelopes artificially increase resulting sub-band bandwidths and cause signal distortion. Results of this incoherent filter do not satisfy properties, such as superposition, that we expect from useful, distortion-free, and predictable filters. Whether the positive and real envelope is found by simple magnitude or Hilbert envelope approaches, the resulting modulation filters are largely ineffective. However, with a more accurate modulation detection foundation, such as with coherent detection as in FM radios, there is the potential to satisfy superposition and other desirable properties in modulation filtering. Well-defined modulation filters are then potentially applicable to a wide range of problems in acoustics, signal processing, and robust speech recognition. Demonstrations of modulation filters include music instrument separation and single-channel multiple talker separation applications. Interesting research questions remain. For example, what should correctly-defined modulation filters do to achieve superposition and how can the need for difficult phase estimates, typically required for coherent detection, be avoided by techniques such as frequency reassignment?
4. Time-Frequency Reassignment and Modulation
Eric Rombokos, Univ. Washington
Abstract:
The spectrogram alone has proven insufficient for approaching the kind of sound perception achieved by nature, and a class of methods called "reassignment" methods which modify the spectrogram may provide a means for improving both time-frequency analysis and methods which model speech as modulation of carriers. I present a brief tour of these methods and issues of their computation and display. Shown are some examples of increased precision in time-frequency visualization, and a discussion of what this means for speech applications. Specifically, under what conditions is finer localization of spectral components useful for tasks which use perceptually-motivated but somewhat broad spectral bands with fixed time windows? Do reassignment estimates of signal components' instantaneous frequencies and local group delays aid in speech recognition, or do confounding factors bury them? For the construction of salient features which really carry the important aspects of speech, it may be more fruitful to use reassignment combined with modulation-of-carrier models.
5. Verbal Query for IP-TV: A Natural Interface to the World of Entertainment
Antonio Moreno, Georgia Tech
Partha Parthasarathy, AT&T Labs
B.H. Juang, Georgia Tech
Jay G. Wilpon, AT&T Labs
Abstract:
The advent of IP-TV enables the user to access an enormous amount of video content (movies, news, TV shows, sports, documentaries, etc.), beyond the traditional TV broadcasting. The challenge in user interface to such a vast selection of information is truly unprecedented (perhaps an order of magnitude more complex than the similar problem associated with the current TV programming guide). Interface by way of natural language query and search is thus imperative. In addition to the traditional text-based search (e.g., google), ASR (Automatic Speech Recognition) represents a fast and friendly retrieval mechanism, particularly in an entertainment environment. In this work, we explore the integration of the speech recognition task with the entry retrieval. In contrast to a conventional ASR 'speech-to-text' engine connected to a 'text-driven retrieval', the strategies pursued aim at bringing the inherent grammatical constraints to the ASR engine to improve accuracy and driving the retrieval system with a richer ASR output.
6. Research and Benchmarking of Advanced Video Search Technologies
Eric Zavesky, Columbia University
Behzad Shahraray, AT&T Labs
Zhu Liu, AT&T Labs
David Gibbon, AT&T Labs
Shih-Fu Chang, Columbia University
Abstract:
The prevalence of content creation and dissemination tools has made digital video a common user experience in many platforms, ranging from Internet, home media centers, to portable devices. To help users cope with the massive video content, new technologies for indexing, searching, and filtering information from distributed video sources are needed. This project involves collaborative study of integrative video search tools leveraging video analysis/indexing research at AT&T and Columbia University.
We are investigating novel methods for combining metadata extracted from audio-visual content with textual transcripts in answering high-level user queries. First, we are developing large-scale statistical models to automatically detect semantic concepts from visual and audio content (e.g., locations, objects, people) and program structures (e.g., news story). We will show how such concepts may greatly improve the video search performance, using a query adaptive algorithm. Second, we will describe on-going work in using AT&T video analysis algorithms (e.g., speaker identification, face localization, and shot segmentation) to analyze style and context of video segments, thereby refining the global search results. Finally, we are exploring the "Force of Many" through mining of concurrent web content (captured by AT&T indexing system over five years from seventeen web/broadcast sources) to discover related salient terms, which are used to augment the video search issued by users. All of the proposed works are benchmarked within the context of TRECVID, a NIST sponsored annual evaluation (currently with more than 70 participating groups) for segmentation, indexing, and retrieval of digital video.
7. Statistical Techniques for Network Data
Eric Vance, Duke University
Tamraparni Dasu, AT&T Labs
David Banks, Duke University
Abstract:
For the AT&T internet backbone we compare patterns of internet traffic between routers and for the same router across time in order to detect anomalous behavior. Instead of making univariate comparisons of data between routers or between times, we use a multivariate distance and depth-based approach to partition the data into bins. These bins allow us to account for dependencies within the data when testing for differences in the underlying structure of the data between routers or over time. We use both multinomial tests and more sophisticated information theoretic techniques to determine whether data is anomalous. Future work includes other depth metrics such as complex hull peeling or simplicial depth.
A second project discussed here is an investigation of telephone calling networks and how these networks change over time. Calls between local AT&T telephone numbers constitute the edges of a graph, and the telephone numbers making them constitute the set of nodes. We use the triad census, shortest-path length, and other summaries of the graph to track and describe changes of the networks over time.
8. A P2P Distributed Storage Service for the AT&T Lightspeed Network
Yih-Farn Robin Chen, AT&T Labs
Jeremy Rahe, University of California, Berkeley
Alan J. Smith, University of California, Berkeley
Yennun Huang, AT&T Labs
Bin Wei, AT&T Labs
Abstract:
AT&T Lightspeed is a high-speed network intended to allow voice, data, and HDTV-quality video-on-demand services to be combined on one network. Providing a large number of HDTV-quality video streams places a large strain on the fiber link connecting each neighborhood to the Lightspeed network. Communication between peer nodes in the same neighborhood, however, is not affected by this restriction. We investigate using a per-neighborhood P2P data storage and retrieval network for video streaming and potentially other uses. Since upload bandwidth at each node is currently limited to 1 Mbps, providing reliable and high speed access to the stored data will require careful placement and replication of data in the network, which is further complicated by transient connection of nodes. We present one possible implementation of such a storage system by using erasure code and a P2P incentive scheme.
Robert R. Miller, AT&T Labs-Research
Saeed S. Ghassemzadeh, AT&T Labs-Research
Vahid Tarokh, Harvard University
Dongwoon Bai, Harvard University
Abstract:
A number of cities have begun using pre-4G WiFi hardware to pioneer broadband "Digital City" access concepts to bridge lower municipal communications costs and provide a source of revenue based on the "Utility Model". These early manifestations have attempted to capitalize upon unlicensed spectrum such as ISM or U-NII bands for outdoor service, leveraging the success systems such as WiFi have had indoors. Although workable for best-effort data traffic, use of unlicensed bands forces the system to contend with interference controlled only by FCC-prescribed device emission limits. Uncontained by building walls, emissions travel further outdoors, increasing interference potential from a variety of sources. As municipal systems mature and traffic grows to become more multimedia rich, interference will eventually limit the ability to provide reliable communications and QoS for streaming services just as citizens and city employees begin to depend on it. A particularly new rich area of research has been the concept of a spectrum "sharing", where a group of users may be "licensed" to use a block of spectrum for designated uses (rather than particular devices operating in "unlicensed" bands ISM and U-NII bands), and where new technologies such as "Cognitive Radio" could be used to mediate spectrum use. In particular, we consider sharing of the spectrum with point to point microwave systems (The primary spectrum users) and neighborhood area networks (the secondary spectrum users). In order to avoid interference, secondary users of the spectrum shall avoid transmission in certain geographic areas around fixed primary transmitter stations. Thus, the secondary users of the spectrum must be able to determine their own position and determine (based on a topographical channel model and using a "shared" database), if they are causing interference to primary users. However, since the wireless communications channels are random, in order to avoid interference, cognitive systems must be either significantly over-engineered (which dramatically decreases their potential capacities) or must be designed using probabilistic approaches.
We propose to formulate the probabilistic design method and analyze the capacity of the cognitive system under both the classic and the probabilistic approaches.
10. Multi-User and Multi-Cellular MIMO Wireless Communication
Arunabha Ghosh, AT&T Labs, Austin
Jeff Andrews, UT Austin
Robert Heath, UT Austin
Runhua Chen, UT Austin
Abstract:
The introduction of multiple transmit and multiple receive antennas (MIMO) has been shown to substantially increase the performance of a wireless communication system. If information regarding the propagation channel is available at the transmitter, "closed-loop" MIMO techniques can more efficiently exploit the benefits of MIMO systems by optimally allocating transmit resource such as power and bits over multiple antennas. A large class of the closed-loop techniques can be characterized in the precoding framework, which includes beamforming, antenna selection or mode selection, etc. These closed-loop techniques significantly outperform the open-loop MIMO techniques where the transmitter in uninformed about the channel. In addition to the single-carrier scenario, multi-carrier precoding has also been studied for future broadband systems such as WiMAX/802.16e. If sufficient number of transmit antennas are available, precoding can be also applied in a multi-user downlink wireless system, notably multi-user precoding and SDMA (spatial division multiple access), to support multiple mobile users at the same time while ensuring no mutual interference.
We have developed novel multi-user precoder design in different transmit / receive antenna settings and demonstrated substantial error performance improvement with transmit selection diversity. Another aspect of MIMO application is in the context cellular networks.
Despite its promise of high spectral efficiency and link quality in a single-user scenario, the application of MIMO techniques in a cellular system is quite a different story. Simulation results have shown that conventional single-user MIMO techniques actually achieve very poor performance in a cellular network, due to the inherent interference-limited nature of any cellular system. Compared to the single-antenna cellular system, co-channel interference in a cellular MIMO system is far more severe because each neighboring base station antenna element acts as a unique interfering source. The capacity and robustness of cellular MIMO systems are severely degraded unless other-cell interference is properly managed. To address this issue, our research has been focused on the interference management in cellular MIMO systems by designing novel space-time power control algorithms. Cellular MIMO power control techniques has been extensively studied and shown to be able to effectively suppress the co-channel interference while guaranteeing a target throughput. Techniques with different levels of channel information requirement and complexity have been developed. Our future research will focus on the networked MIMO communication system with cooperatively scheduled transmission, where base stations collaborate with each other to reduce co-channel interference, allowing for a network-level interference management solution.
11. Failure management in Commercial Data Centers
Kaustubh Joshi, UIUC
William Sanders, UIUC
Matti Hiltunen, AT&T
Richard Schlichting, AT&T
Abstract:
Dealing with failures in large data center environments is a challenging activity that requires close operator supervision. However, the dynamic nature of application configurations and workloads makes dealing with and reasoning about the large number of alarms generated by current monitoring systems challenging. In this talk, we will describe our ongoing research efforts to couple measurements of correlated application agnostic parameters such as processor loads together with models of system topology in order to provide failure detection and isolation that is robust across changes in system load. The developed monitors can be used in isolation or in conjunction with a framework for probabilistic diagnosis and recovery we have developed in order to provide semi-automatic failure management capabilities to data-center operators.
12. Towards a Free-Viewpoint Video System
Alan O'Connor, Harvard University
Shankar Krishnan, AT&T Labs
Amy Reibman, AT&T Labs
Vinay A. Vaishampayan, AT&T Labs
Roger Brockett, Harvard University
Abstract:
This summer project explored techniques for creating free-viewpoint video. This technology allows a user to watch a real scene from an arbitrary point-of-view by combining views captured from a small number of cameras. This is an active area of research both in the computer graphics and the video processing communities.
After learning about camera models, projective geometry, and recent work on free-viewpoint video, I created a new system using the image-based visual hull approach. This system consists of several parts which perform the following tasks: Camera calibration, foreground/background segmentation, computation of the new depth field, and texture mapping from the reference images to the synthetic view. I will describe the algorithms that underlie this system and will describe some of the issues involved in extending this work to a system that operates on live video sequences.
13. Next Generation Video Coding and Image/Video Quality Assessment Algorithms
Pierre Costa, ATT Labs, Austin
Prof. Al Bovik, UT Austin
Kalpana Seshadrinathan, UT Austin
Sumohana Channappayya, UT Austin
Abstract:
In this talk we present past and ongoing collaborative research efforts at ATT Labs Austin and the University of Texas at Austin. First, we present an algorithm that helps convert variable bitrate (VBR) MPEG2 streams into constant bit rate (CBR) data for better bandwidth usage on CBR networks such as ATM networks. This algorithm encapsulates the data packets in RTP headers and provides feedback using RTCP. We then present the contributions made to full reference image and video quality assessment. The Structural SIMilarity (SSIM) index and the Visual Information Fidelity (VIF) measure of image quality are presented. Both these measures have been very widely accepted by the image processing community and are on their way to becoming the de-facto image quality measures (if not already there). We then present ongoing research on video quality assessment and conclude by mentioning potential applications for these new quality measures.
14. Incorporating Prosody in Speech Translation
Srinivas Bangalore, AT&T Labs
Vivek Rangarajan, University of Southern California
Shrikanth Narayanan, University of Southern California
Abstract:
Prosody refers to the rhythm and intonation patterns of spoken language that convey meaningful information beyond the orthographic transcription like emphasis, intent, attitude and emotion of a speaker. Our work is motivated by the desire to incorporate prosody within a speech-to-speech translation framework. Typically, state-of-the-art speech translation systems have a source language recognizer followed by a translator. The translated text is then synthesized in the target language with prosody predicted from text. In this process, the prosodic information present in the source signal is lost during translation. However, with reliable prosody labeling in the source language, the prosody can be transferred to the target language (e.g., English-to-Spanish) and the predicted prosody can used by a TTS system in synthesizing speech with appropriate prosody. We describe our maximum entropy modeling framework for modeling the syntactic stream and a traditional Hidden Markov Model (HMM) to model the acoustic-prosodic stream for automatic prosody labeling of the source language.
15. Using Tree-Drawing Algorithms to Visualize Tree-Like Networks
Yehuda Koren, AT&T Labs
Adrian Rusu, Rowan University
Abstract:
As a networking company, AT&T developed advanced tools for network visualization, which include the Graphviz graph drawing package. These tools facilitate direct examination of the network that can lead to better understanding of its structure. Many of the networks generated in AT&T are trees or tree-like. That is, they contain significant cycle-free components. Since trees are both simple and ubiquitous, a lot of research has been done on visualizing trees, which has produced a plethora of tree-drawing algorithms. We present some of the most important tree-drawing approaches and propose to apply them in the visualization of tree-like networks.
16. XTreeNet: A Scalable Unified Overlay Network for XML Content Access and Distribution
K.K. Ramakrishnan, AT&T Labs
Divesh Srivastava, AT&T Labs
Michael Rabinovich, Case Western Reserve University
Yin Zhang, University of Texas at Austin
Jon Gibson, University of Texas at Austin
Emiran Curtmola, UC San Diego
Abstract:
Large amounts of data are being created and distributed using eXtensible Markup Language (XML). XTreeNet is an XML aware overlay network designed to efficiently support two of the major content access models - publish/subscribe and query/response - in a unified system. XTreeNet uses the concept of content descriptors (CDs) and core-based CD distribution trees to provide fast routing within the overlay - providing dynamic content-based distribution with the lower overheads normally found in topic-based solutions. We present an analysis of how CDs are generated in real-world documents including online databases such as Wikipedia and content publication systems such as RSS. We show how, even in a difficult case such as a free-form database, the number of CDs generated presents a feasible space in which to work. Additionally, we evaluate XTreeNet's performance with extensive simulations using service times garnered from a XTreeNet prototype.
17. Distributed Route Control for Customer Traffic Load-Balancing
Patrick Verkaik, UCSD
Kobus van der Merwe, AT&T Labs
Dan Pei, AT&T Labs
Aman Shaikh, AT&T Labs
Alex C. Snoeren, UCSD
Abstract:
To improve the reliability of Internet access, customers and data centres often connect to AT&T using multiple access links. However, in many cases customers find that inbound traffic usage of their access links is heavily unbalanced and underutilises the available access link capacity. In order to aid customers in addressing this problem we propose a solution that allows an ISP to flexibly engineer interdomain routing on a per-customer, per-router basis, taking input from both network-wide and customer-specific policies. Our solution is based on a recent result of AT&T Research: the Intelligent Route Service Control Point (IRSCP) architecture. IRCSP implements precise and manageable control of interdomain routing, presenting a powerful interface that enables a large range of traffic-engineering applications of which customer traffic load-balancing is only one example. IRSCP is currently deployed by AT&T but does not yet handle applications of this scale. In this talk we focus on our work in developing a scalable, robust IRSCP. We propose to scale IRSCP to control thousands of interdomain routers and discuss our solutions to challenges of achieving scalability and fault-tolerance while maintaining correctness of the IRSCP.
18. Emulating Large Scale Enterprise Networks
Calvin Hubble, UCSD
Rick Schlichting, AT&T Labs
Jim Pelletier, AT&T Labs
Amin Vahdat, UCSD
Abstract:
The networks of AT&T's large enterprise customers are becoming increasingly complex and heterogeneous, with multiple locations scattered across the globe, a wide collection of different networking technologies, and devices that can range from high-end servers to RFID and other small devices. Understanding the behavior of these networks and how they impact higher-level services is challenging and exacerbated by the lack of a platform supporting realistic experiments. This talk will give an overview of how emulation technology based on the ModelNet system is being used as the basis for building an experimental platform that can model large scale enterprise networks.
19. Simulation and Analysis of Failures and Impairments for Multimedia Distribution Networks
Justin Burke, UCSD
K.K. Ramakrishnan, AT&T Labs
Robert Doverspike, AT&T Labs
Dongmei Wang, AT&T Labs
Guangzhi Li, AT&T Labs
Kostas Oikonomou, AT&T Labs
Amin Vahdat, UCSD
Abstract:
Large carriers are now offering or planning to offer video services over an IP infrastructure (e.g., broadcast TV, Video-on-demand). To provide a transport network that is cost-competitive with broadcast cable, there is keen interest in using IP layer Protocol Independent Multicast (PIM). Furthermore, these services need to preserve an objective degree of multimedia QoS under potential transient network faults. To mitigate this, various forms of MPLS or lower layer "Fast reroute" can be used to address the most common single network failures. Subsequent interaction with OSPF reconfiguration and PIM routing tree reconfiguration is an area of study. We have developed a simulation tool for examining the system architecture, fault tolerance mechanisms, and network reconfiguration latencies of such a system under all possible failure scenarios. We will present a high level description of our tool and approaches for this problem. To protect proprietary information, specific network results will not be presented.
20. Toward a Fair and Robust Internet Backbone Routing
Sen Shubho, AT&T Labs
Oliver Spatscheck, AT&T Labs
Bill Lin, UCSD
Jerry Chou, UCSD
Abstract:
As the new technology and applications emerge in the network, the behavior of traffic on the Internet is dramatically changing. The traditional static routing like OSPF is no longer sufficient to handle the variable traffic. Our goal is to maximize the network throughput with minimum network resource, guarantee the routing performance under variable traffic demand and insure the fairness between flows. In this study, we compare several routing algorithms and further explore a new one. More specifically, we formulate the problem as a maximum concurrent flow optimization problem to exploit path diversity and load balance traffic across network to sustain the changing traffic with proportional fairness rate assignment.
21. Multi-Level RAID for Very Large Disk Arrays - VLDAs
Alexander Thomasian, New Jersey Institute of
Technology
alexthomasian@gmail.com
Abstract:
Very Large Disk Arrays - VLDAs constitute a new paradigm for cost-effective and ultrareliable storage of high volumes of data. VLDAs use storage nodes - SNs (bricks) as building blocks, where each SN has a RAID controller, cache, and a dozen disks, which may be a k-disk-failure-tolerant - kDFT array. A Multilevel RAID paradigm based on mirroring or erasure coding at the higher level is required to cope with SN failures, e.g., RAID6/5 is based on a RAID6 paradigm across SNs and RAID5 across disks, so that two SN and one disk failure per node can be tolerated. We discuss the layout of parities for MRAID and propose storage transactions to ensure the atomicity of storage updates. Performance and reliability analysis of SNs are areas of further investigation.
22. Troubleshooting in MPLS Virtual Private Networks
Z. Morley Mao, University of Michigan
Dan Pei, AT&T Labs
Jia Wang, AT&T Labs
Ying Zhang, University of Michigan
Abstract:
Multi-protocol label switching (MPLS) Virtual Private Networks (VPN) has had significant and growing commercial deployments, and is a very important service provided by AT&T. VPNs usually carry applications, such as VoIP, data replication, and financial transactions that are sensitive to possible performance degradation due to routing failures. As a result, network operators need to identify the failure and react quickly, but unfortunately, there is no existing tool that can help the operators do that.
The biggest challenges in troubleshooting the network are the network scale (e.g. there are 800K routes in AT&T VPN networks, and thousands of routers and interfaces) and the sheer volume of the measurement available data (millions of BGP, syslog, and OSPF messages per day).
In this talk, we present our initial troubleshooting analysis that aims at converting large volume of various measurement data into a few dozen actionable reports about significant routing disruptions via a series of systematic approaches. Given VPN's distinct hub-and-spoke structural characteristic, it is important to troubleshoot the failures which impact the popular hubs. This troubleshooting tool, once fully developed and deployed, can greatly help network operators react to failures quickly prioritized by the severity of the events and facilitate AT&T to improve its VPN network performance significantly.
23. Technology Diffusion and Long Term Forecasting: Application to Growth of Wireless Mobile Services
M. Hosein Fallah, Stevens Institute of
Technology
Elias Aravantinos, Stevens Institute of Technology
Abstract:
Demand for wireless services have been growing much faster than anything we have experienced in the past. The level of deployment of wireless mobile, however, varies widely from cities to countries to regions of the world. The classical technological forecasting models such as Gompertz, Fisher-Pry and others generally fit a logistic function (S curve) to the historical data to project future growth. These models do extremely well for short term forecasting where the environment is assumed to be static. Dynamic changes in the user environment due to competing technologies such as Broadband make any prediction beyond a year or two suspect. Longer term prediction is affected by other factors such as demographics, technology evolution, economics and culture. We are studying longer term technology diffusion forecasting by applying moderating factors to the classical models using analogy and growth observed in the leading markets. We will examine growth of the wireless mobile in US and leading European countries using Gompertz diffusion model and illustrate opportunities for improving the forecasting model by applying an analogy approach.
24. Robust and Loss-Tolerant Link and Transport Protocols for Wireless Network Environments
Vijay Subramanian, RPI
K.K. Ramakrishnan, AT&T Labs
Shiv Kalyanaraman, RPI
Abstract:
The widespread deployment of wireless links and the expected explosion in the use of wireless broadband systems makes it important for link and transport layer protocols to perform well in these environments. Wireless paths, especially multi-hop paths present challenges in terms of bounded end-end delay, end-end loss rate experienced by the transport layer and achievable goodput. Current transport protocols such as TCP-SACK are known to fail as the error rate goes above 5 % PER. Moreover, link-level protocols interfere with transport layer mechanisms due to large number of ARQ attempts causing high and variable latency. We develop transport and link-level protocols that make use of the general principles of fragmentation, loss estimation and Forward Error Correction (FEC). Our design structures these building blocks to perform well even under high loss rates (up to 50 % PER). Analysis of the scheme enables us to tune the link and transport protocols based on their different needs while providing insight into the optimal configuration. Moreover, we illuminate the issue of balance of functionality and interactions between the link and transport layers. At the link layer, these mechanisms are designed to meet the objectives of having bounded delay while exporting a negligible residual error rate. At the transport layer, we concentrate on tackling the residual error rate while operating on a longer time-scale. Our proposed mechanisms make it possible to operate efficiently under conditions of high loss rates including highly bursty environments.
Networks Lab, RPI
http://poisson.ecse.rpi.edu/~vijay
25. Underwater Sensor Networks: Applications and Challenges
Reda Ammar, Computer Science & Engineering,
UCONN
Jun-Hong Cui, Computer Science & Engineering, UCONN
Zhijie Shi, Computer Science & Engineering, UCONN
Thomas Torgersen, Marine Sciences, UCONN
Shengli Zhou, Electrical & Computer Engineering, UCONN
Abstract:
The Earth is a water planet. For decades, there have been significant interests in monitoring aquatic environments for scientific exploration, commercial exploitation and coastline protection. Highly precise, real-time, and temporal-spatial continuous aquatic environment Monitoring systems are extremely important for various applications, such as oceanographic data collection, pollution detection, and marine surveillance. However, traditional techniques, such as remote telemetry and sequential local sensing, cannot satisfy these high-demanding application requirements.
Recently, sensor network has emerged as a very powerful technology for many applications, including monitoring, measurement, surveillance and control. The idea of applying sensor networks in underwater environments (i.e., forming underwater sensor networks) has received increasing interests. Even though underwater sensor networks (UWSNs) share some common characteristics with terrestrial sensor networks, such as the large number of nodes and limited energy, UWSNs are significantly different from terrestrial sensor networks in many aspects: low bandwidth capacity, large propagation latency, node float mobility (resulting in high network dynamics), high error probability, and 3-dimensional space. These new features bring many challenges to the design of UWSNs.
In this talk, I will review the unique features of UWSNs and discuss the research issues in UWSN design. Adopting a top-down approach along the layered protocol stack, we will roughly go down from the top application layer to the bottom physical layer. At each layer, a set of new design challenges will be identified.
* Please note that this is a team effort. Only representative faculty members in different fields are present above. For a complete list, please refer to the website of the underwater sensor network (UWSN) lab at UCONN, http://uwsn.engr.uconn.edu
26. Experiments and Performance of RFID Systems
C. G. Wang, University of Arkansas
K. Sohraby, University of Arkansas
M. Daneshmand, AT&T Labs
Abstract:
With the support of AT&T RFID Lab during the summer 2006, we conducted experiments on RFID Gen-1 and Gen-2 tags. It was found that there are many factors which influence the communication between readers and tags, including antenna power and direction, space between tags, tag moving speed, object materials, etc. Even though Gen-2 protocol achieves improved performance and higher flexibility than Gen-1 protocol, Gen-2 protocol demonstrates a level of unreliable reading in our experiments. However there is not quantitative analysis of the performance of Gen-2 protocol in the current literature. We established a discrete-time Markov chain (DTMC) model to quantitatively analyze Gen-2 protocol and calculated successful tag identification rate (STIR) and tag identification speed (TIS). This model was effectively verified by extensive simulations. One of the important properties of Gen-2 protocol due to the Q algorithm is that its TIS is nearly independent of the number of total tags. This algorithm reduces and resolves tag-collisions more efficiently and effectively than Tree traverse algorithm used in Gen-1 protocol. Our quantitative analysis helps evaluate Gen-2 system and design enhanced algorithm to further improve the performance of Gen-2.
27. Reliability in RFID systems
Ahmad Rahmati, Rice University
Matti Hiltunen, AT&T
Rittwik Jana, AT&T
Lin Zhong, Rice University
Abstract:
Radio Frequency Identification (RFID) systems are becoming increasingly popular in applications such as access control, location tracking, industrial automation, and as a replacement for barcodes in consumer products. RFID systems may use either active or passive tags. Active tags have an onboard power source (battery) and they can support complex operations and protocols, but are impractical for many applications because of cost, power, and size issues. Passive tags do not have any onboard power source and rely on the energy from the reader's RF signal to operate. They are more economical, but severely limited in functionality. With the decreasing price trend of passive tags, there remain two main challenges that prevent large-scale deployment of RFID systems; namely privacy and reliability. Here, we are studying the reliability characteristics of current state of the art 'Gen 2' RFID systems and researching reliability methods to address this issue. We are also implementing a passive RFID object/person tracking system as a test bed for our research.
Ahmad Rahmati,Graduate Student,
Rice Efficient Computing Group, Dept. of ECE, Rice University, Houston, TX
http://www.ruf.rice.edu/~rahmati/
M. Hosein Fallah, Stevens Institute of
Technology
Jiang He, Stevens Institute of Technology
Abstract:
Although knowledge spillovers are considered critical to the speed of innovation particularly in technology clusters, the relationship between characteristics of knowledge spillover network and development of regional clusters is not well understood. Based on patent co-authorship data, we are studying the innovation networks in communications technology clusters. We are analyzing these networks for two geographical telecom clusters - New Jersey and Texas. These networks have evolved longitudinally as the clusters were undergoing different stage of lifecycles. We explore the similarities and differences in dynamics of the innovator networks for the two clusters over different time period. Understanding this dynamics will help the industry and regional planers implement strategies to increase innovation output and drive the industry growth in the region.
29. Optimizing Large File Transfers with WORMHOLE
Hasan Abbasi, Georgia Tech
Prof. Karsten Schwan, Georgia Tech
Matti Hiltunen, AT&T
Rick Schlichting, AT&T
Vinay Vaishampayan, AT&T
Abstract:
With increased capacity for storage and computation there has been a corresponding increase in the size of data files for both enterprise and scientific applications. The cost of transferring the generated data has become astronomical as data file sizes have dwarfed available network bandwidth. We propose a framework for reducing file transfer time over the internet for file sizes ranging from 1GB to 100GB. Such file sizes can be observed in video editing applications, semiconductor design applications, and scientific simulations. Additionally we have observed that these applications follow a pattern in transferring the files over a WAN, our approach takes advantage of this to further optimize the file transfer. To reduce the latency of file transfer, we utilize three main methods: delta compression, file staging in the core network, and predictive transfer. Using a combination of these techniques, we hope to reduce the file transfer time to more manageable levels. We evaluate our framework using a combination of scientific data and randomly generated/edited files.
30. Better Traffic Matrix Estimation in a Large ISP Network
Qi Zhao, GA Tech
Abhishek Kumar, GA Tech
Jun Xu, GA Tech
Zihui Ge, AT&T Labs
Jia Wang, AT&T Labs
Abstract:
The traffic volume between origin/destination (OD) pairs in a network, know as traffic matrix, is essential for a number of network management and tasks in operational IP networks such as capacity planning and traffic engineering. In the collaboration with AT&T research labs, we design two novel methods to produce better traffic matrix estimation. The first proposes a brand-new data streaming algorithm to directly measure traffic volume for each OD pair. It makes each ingress node and egress node generate traffic digests that are orders of magnitude smaller than the raw traffic stream. By correlating the digests collected at any OD pair using Bayesian statistics, the volume of traffic flowing between the OD pair can be accurately determined. It achieves around one order of magnitude higher estimation accuracy than the network tomography approach (including Tomogravity method) and is multiple times better than the sampling scheme (e.g., NetFlow) given the same amount of data generated. The other method combines the SNMP link counts and sampled NetFlow records together to produce more accurate estimation of traffic matrices even when NetFlow records are available on only a subset of ingress nodes. In this work we also design methods that, by comparing notes between the link counts and flow records, identify and remove dirty data (measurement errors in SNMP and NetFlow due to hardware, software or transmission faults). These proposed methods not only improve the accuracy of traffic matrix estimation, but may also benefit a number of other applications that depend on these data.
We also believe that there exist some implicit temporal correlations among OD flows at different time instants could help increase the estimation accuracy further as another information source. This is a potential rich research area that has barely been explored. We are investigating to find a good underlying model which captures both temporal and spatial correlations among OD flows. Based on this model, some statistical methods could be carefully designed to infer the traffic matrices more accurately given the existing sampled NetFlow records and SNMP link counts. Having this model in hands we also expect to improve the performance of dirty data removal significantly. More interestingly this work would enable us to do traffic matrix prediction. A discrepancy between the prediction and what really happened could indicate that an unusual change (anomaly) appears somewhere.
31. Toward Fully Automated Verification and Optimization of the RUBY System
Min Gyung Kang, CMU
Zihui Ge, AT&T Labs
Jia Wang, AT&T Labs
Abstract:
RUBY (RUle Building capabilitY) is a platform that AT&T developed and used for dynamic generation and management of expert rules that automates critical operations such as customer care, network care and billing dispute. Nowadays, RUBY processes hundreds of thousands network alarms, orders and tickets each day. The RUBY system, which adopts the object-oriented design, is inevitably complicated due to the scale of the system and the dependency among the rules. Since most of the rules are created manually by different "rule managers" across different domains, it is crucial to verify the system-wide integrity and to ensure high performance of RUBY in a systematic manner. In this talk, we present our framework on automatic verification and optimization of RUBY. Our analysis is based on rules configuration and system log. We present our results from a preliminary examination of the current rules using the prototype of our checking tool. In addition, we suggest that improving this tool is promising to help the rule managers and to enhance the business processes.
32. Use of Communication Graphs for Real-time Anomaly Detection
Rangarajan Vasudevan, University of Michigan, Ann
Arbor
Z. Morley Mao, University of Michigan, Ann Arbor
Kobus van der Merwe, AT&T Labs
Oliver Spatscheck, AT&T Labs
Abstract:
Today's ISPs, from their vantage point as operators of the core network, are in a position to detect the presence of many anomalies. Some of these like DDoS attacks have plagued the networking community for some time now while no satisfactory solution exists for their detection. Others like phishing attacks for sensitive information have traditionally been countered at the end-host level, but are not effective against zero-day and new attack types.
We propose a tool that adopts a communication graph-based approach that scales to hundreds of thousands of customers, and demonstrates potential to detect many anomalies like DDoS attacks, botnet attacks, phishing attacks, spam, worm propagation etc. The approach involves monitoring communication and traffic patterns across time and space. Across time monitoring helps identify anomalous traffic patterns as deviations from the norm. Across space monitoring helps identify network-wide anomalies that plague multiple customers. Our basic algorithm involves building communication graphs from observed communication patters. Then, we use heuristics to represent anomalies as properties in the graph. Thereafter, the problem of anomaly detection essentially becomes a problem of discovering specific properties in these graphs.
33. Anticipating Customer Behavior through Traffic Analysis
Manfred Georg, Washington University in St.
Louis
Kadangode K. Ramakrishnan, AT&T Labs
Nicholas G. Duffield, AT&T Labs
Abstract:
We are currently analyzing VPN (Virtual Private Network) packet traces collected using Gigascope. Using this data we model the behavior of customers and seek an efficient grouping which minimizes the aggregated capacity that needs to be allocated for these customers while meeting their SLAs. The motivation for this mainly stems from the desire to migrate such customers to newer shared packet based access networks. We explore different methods for classifying the customers according to their traffic characteristics and correlate simple identifiers such as contracted CIR and number of unique IP addresses with traffic characteristics. The intent is also to understand the weaknesses of the use of coarse grained average statistics for capacity management, as opposed to the use of fine-grained packet trace information.
34. Darkstar - Anomaly Detection
Juan Caballero - CMU
Jennifer Yates - AT&T Labs
Ajay Mahimkar - University of Texas, Austin
Zihui Ge - AT&T Labs
Dawn Song - CMU
Shobha Venkataraman - CMU
Jia Wang - AT&T Labs
Abstract:
Automatic identification of anomalies on network data is a problem of fundamental interest to network operators that want to diagnose incipient problems in their networks.
Network operators gather multiple and diverse data sources from the network for monitoring, diagnostics or provisioning tasks. Finding anomalies on this data is a huge challenge due to the volume of the data collected, the number of data sources and the diversity among those data sources.
Previous approaches for automatic anomaly detection have focused on a single data source or small sets of traffic features. Our work is a step towards generalizing automatic anomaly detection to multiple data sources and traffic features.