Visualizing Large-Scale TelecommunicationNetworks and Services

James AbelloEmden R. Gansner

Daniel KeimEleftherios Koutsofios

Stephen C. NorthRuss Truscott

Network Services Research Labfabello,erg,keim,ek,north,truscottg@research.att.com

http://infolab.research.att.com

September 15, 1999

1 Introduction

Global telecommunication networks and services areamong the enterprises having the highest volumesof real-time data. A voice network may completemore than 250 million calls per day. Each is de-scribed by one or more events, yielding a total of tensof gigabytes of data daily. Wireless, AsynchronousTransfer Mode (ATM), frame relay, Internet Protocol(IP) networks and higher-level services on them alsoare described by massive data sets, and can presentadditional problems in reconstructing an end-to-endview of user activity. Understanding this data atfull scale is crucial for managing networks and im-proving their performance and reliability from a cus-tomer’s viewpoint. Visualization techniques have be-come increasingly important to achieving this goal.

In the AT&T Infolab, we have developed a new visu-alization platform for the interactive data explorationof large networks. The platform incorporates inter-

active 3D maps and diagrams, statistical displays,network topology diagrams, and pixel-oriented dis-plays, while supporting a variety of display technol-ogy, including a wall-sized screen.

Its applications include monitoring and analyzing ac-tivities at the network element, network-wide, cus-tomer and service levels. These activities maybe network generated (e.g., exploration of networkevents and alarms or customer generated (e.g., usageanomalies such as fraud). End uses of the analysiswould likely include improvement of service to cus-tomers, market analysis and gaining an understand-ing of previously hidden relationships between andwithin data segments.

In this paper, after an initial description of the typesof applications the platform is aimed at, we con-centrate on three novel aspects of the system. Wedescribe our use of large displays as one approachto handling large data sets. We then describe the

1

SWIFT-3D network viewer, which provides the mainsoftware base for the platform. Finally, we considera new visual metaphor for viewing large networks.

2 Telecommunication Applications

The goal of this work is to support interactive vi-sual exploration of databases that describe full-scalecommercial telecommunication networks, and to si-multaneously raise the level of abstraction in visu-alization, for example showing layered services ornetwork performance from an individual customer’sviewpoint. Derived from this goal is the ability tomove from data to business decision within minutes.

Many data analysis tasks that are tractable on smallor medium-sized data sets can be difficult at greaterscale. When practitioners refer to terabyte databases,they sometimes mean databases of image, sound orvideo data. In contrast, our application involvesworking with many small records describing trans-actions and network status events. The data process-ing involved is different in terms of the number ofrecords and data items to be interpreted. In voicenetworks, the detail record for each call conforms toan industry standard format (Automatic Message Ac-counting, or AMA) that has about 50 attributes suchas originating and terminating phone numbers, date,time and duration of the call. In our application thisinformation is stored for each of the hundreds of mil-lions of calls made daily, yielding about 15 GByteof data uncompressed. In addition, data is collectedfrom the other networks previously mentioned. Un-derstanding the relationships between them is in-creasingly important, e.g. to manage integrated com-munication services for global enterprises, but thedata management problems that result are even morechallenging than for a single service.

More than just scale is involved: our goal is also toraise the level of abstraction in network visualiza-tion, and to improve the real-time response of ouranalyses. This can help network managers and busi-ness decision makers to recognize and respond tochanging conditions quickly; within minutes whenpossible. A scalable research prototype for visuallyexploring full-scale network traffic must therefore

provide good interactive response, avoid instance-specific processing, and be flexible enough to sup-port experiments in both back-end queries and theuser interface. In our initial experiments, we foundthat commercial database systems either couldn’thandle such large volumes or consumed far toomany resources. Another problem with commercialdatabases is that the administrative effort was toohigh to support experimental research. Databases,however, have many useful features, such as dataindependence, and a standard query language. Thetools we have built have some of these features. Amain difference in our approach is the emphasis ondata streaming, in comparison to a query/responsemethodology employed by formal databases.

3 Large Displays

One approach to the scale problem is to use a phys-ically large display. Figure 1 shows our displaywall, inspired by projects such as the CAVE[7] andPowerwall[8].

Figure 1: The Infolab Wall

Our main display wall (60� 15

0) is driven by 8 LCDprojectors. These are connected through a software-controlled video switch, usually to display the out-put of two graphics pipes of an SGI Onyx. The sameOnyx can drive a smaller (70

� 90) 4-projector wall

elsewhere in the same building, using its third pipe.Other compute and disk servers for network dataanalysis projects are connected on an 800-megabit

2

High Performance Parallel Interface (HIPPI) net-work, providing 10 terabytes of on-line storage andanother 20 terabytes of tape under hierarchical stor-age management.

4 Visualization in SWIFT-3D

The principal application that runs on the displaywall is SWIFT-3D, an interactive, large-scale net-work viewer [10]. The SWIFT-3D system inte-grates a collection of relevant visualization tech-niques ranging from familiar statistical displays, topixel-oriented overviews with interactive 3D-mapsand drag+drop query tools. It provides comprehen-sive support for data exploration, integrating largescale data visualization with querying, browsing andstatistical evaluation (see [3, 4, 5, 11] for examplesof previous related work). The visualization compo-nent maps the data to a set of linked 2D and 3D views[6, 12] created by different visualization techniques:

� Statistical 2D Visualizations (line graphs, his-tograms, etc.) - used as overview displays andfor interactive data selection

� Pixel-oriented 2D Visualizations - intended asbird’s-eye overviews and for navigation in 3Ddisplays

� Dynamic 3D Visualizations - used for an inter-active detailed viewing of the data from differ-ent perspectives.

In addition, the system provides tightly integratedbrowsing and querying tools to select the data to bedisplayed and to drill-down for details if some inter-esting pattern has been found.

A screenshot of the overall system is given in Fig-ure 2. The upper left window shows a time line visu-alization of voice network volume in 10-minute in-tervals. This plot shows the volume for different ser-vices (e.g. residential, business, and 1+ dial-aroundservice, software-defined networks, and aggregatevolume). The window below the time line allowsthe user to select data for display by date, time or

type of service. The large window shows a three-dimensional display of the data using a histogramspike for each location to display a value (typically,level of activity) corresponding to the cursor positionin the time line window (11:00). The user can inter-actively navigate in the 3D display, zoom in at inter-esting locations, or view the map from arbitrary per-spectives. An automated path-planning module hasalso been designed to determine a natural, context-preserving path from one viewpoint to another. Themapping between the data and display objects is setin an auxiliary file that contains geometric informa-tion about points, lines, polygons, and triangles, andcoloring. Various color maps may be defined to high-light interesting properties of data. The mapping filemay contain multiple levels of detail; for example,a data set representing the United States may be di-vided according to state, county, and telephone ex-change, census block and 9-digit postal zip code out-lines. Also, multiple data value sets can be mappedto the same geometry. For example, we can map statepopulation to the state outline level and county pop-ulation to the county level. As the view of a state en-larges, the displays can shift from showing a singlevalue for state population to showing one per county.The user may also play through an adjustable inter-val in the time line window to get an animated time-sequence display (see video). If the user sees an in-teresting pattern in the visualization window, a drag-and-drop interface is available to drill-down to getdetails, explore context and take actions if necessary.This provides an intuitive way of converting spatialinformation into detailed information such as the toporiginating or top dialed numbers.

An additional 2D overview window is provided,showing call volume for each location by one col-ored pixel (cf. lower left corner of Figure 2). Thetechnique behind the pixel-oriented 2D overviews isan adaptation of the Gridfit approach used in the Vi-sualPoints system [9]. Gridfit places data points ona pixelated display, so that points having coordinatesthat would normally map to the same display pixelare represented by other pixels that would otherwisebe unoccupied. Its algorithm is based on hierarchi-cal partitioning of the data space, using a top-downreallocation of the screen space according to the re-quirements of subregions. Gridfit allows an efficient

3

and effective repositioningof the pixels on the screensuch that the (absolute and relative) position of thedata points and their distance is preserved as muchas possible. The color is chosen such that high callvolumes are mapped to dark colors and low call vol-umes are mapped to bright colors.

Figure 2: SWIFT Overview Screen

Figure 3: SWIFT Time Series

4.1 Implementation Architecture

SWIFT-3D consists of three modules: data collec-tor, aggregator, and visualization interface. Thesecommunicate using self-describing data-independentbinary formats consisting of a header that definesrecord size, type, and data context, followed by theactual data. This is necessary since SWIFT-3D isdesigned to work in real-time: the data processingmodules can work incrementally and the visualiza-

tion tools can safely access the data files while theyare being updated.

To achieve good performance, SWIFT-3D uses vari-ous techniques that minimize processing delays andthe use of system resources. Such techniques includeprocessing pipelines, direct IO, memory mapping,and dynamic linking of on-the-fly generated code.

4.1.1 Data Collection and Storage

SWIFT-3D must collect data from many differ-ent sources having their own specialized formats.SWIFT-3D includes tools to convert such data to itsinternal self-describing format. When data is alreadyin a fixed format, all that needed is to associate adata record schema with the file. In most cases, stan-dard tools suffice for data conversion, but some typesof data need more intricate pre-processing. For ex-ample, the voice network generates records in thepreviously-mentioned AMA format. This format hasmany sub-record types that can be combined to de-scribe a call. Extracting information from AMA filesis further complicated because, depending on thetype of call, a value can be stored in different sub-records. For example, the dialed number is kept indifferent places in domestic and international calls.Such idiosyncratic processing is performed by cus-tom tools to load into SWIFT-3D format.

4.1.2 Data Processing

Initial processing of a data feed usually involvesreading in records and computing basic statistics.SWIFT-3D relies on a stream pipeline model. Ac-cessing large-scale data on disk can be expensive,so instead of storing the output of each processingstep to disk, the stream processors are implementedas concurrent processes that exchange data. SWIFT-3D extends the UNIX pipe model of single writer andsingle reader to that of single writer and many read-ers to minimize data copying. For example, we mightneed to process a day’s worth of telephony data andcompute: (1) how many calls there were per areacode and exchange (NPA/NXX), (2) how many callsdid not complete and separate these by failure type.This could be implemented by a single process that

4

reads data from disk and feeds two other processesto count.

SWIFT-3D provides several tools to be used in suchpipelines. These include tools to filter records (e.g.remove calls that did not complete), count based onattributes (e.g. count number of incomplete callsby NPA/NXX), split a single file into several basedagain on some combination of attributes (e.g. sepa-rate calls into a file per type of service such as tollfree calls, operator calls, collect calls, etc.).

The expressions used for filtering, counting, andsplitting are specified as C-style expressions. For ex-ample, the expression

if (tos == TOLLFREE && iscomplete) KEEP;else DROP;

filters out calls that are not toll free (1-800, 1-888, 1-877) or not complete. These expressions are turnedinto C code that is compiled into shared objects onthe fly and are then dynamically linked in and exe-cuted. This approach combines the speed of com-piled code with the flexibility of tools such as AWK.C and AWK seem to be the tools of choice for mostof our statisticians and analysts.

4.1.3 Data Visualization

SWIFT-3D’s visualization tools allow users to ex-plore data filtered by the stream processors. Theyare designed to be interactive in the sense that theuser can view some data set, focus on a specific sub-set, query the system for the raw data that gener-ated this subset, re-aggregate and view the new re-sult. SWIFT-3D enables this by keeping enough in-formation to link raw data, aggregate data and visualobjects. This link to visual objects is implementedby generating geometric data sets that contain infor-mation about the items they represent. For exam-ple, an NPA/NXX may be represented by a point,bar, or polygon of its geographic area. In all cases,the geometry file contains information to link theNPA/NXX to the point, line, or polygon. Besidesanswering user queries, this facility is also used toalter the geometry based on data values. For exam-ple, if NPA/NXXs are shown as polygons and busy

NPA/NXXs need to be colored red, the system usesthis mapping to determine red polygons.

For reading records off disks, SWIFT-3D usesDirect-IO if available. Direct-IO bypasses kernelbuffer copying from disk, and can be twice as fastas normal IO. (Normal IO can be faster for data thatwas recently read and is still in cache, but given thesize of our datasets, this is rarely the case).

The format of the counts files is also self-describing.Such files implement a type of two dimensional arrayof values (integers, floats, etc.). One dimension (the‘frame’) corresponds to time buckets while the other(the ‘item’) corresponds to the aggregation type. Thesecond dimension can be accessed using a dictionarythat maps item ids to item positions. For example,the second dimension in the scenario above may havean entry per NPA/NXX observed, and the dictionarymight indicate that data for NPA/NXX 973-360 is inposition #0.

These files are designed to be accessed and changedincrementally: when new data arrives, these files areopened and the various counts are increased in place(using some buffering to minimize accesses to thefiles). The actual updating of the files is done us-ing memory mapping, due to the random access na-ture of the updating. File locking is used to protectagainst accessing such a file in the middle of an up-date. Also, each update increments a count stored inthe file. This allows the visualization tools to effi-ciently check if the file has been modified.

4.2 SWIFT-3D Applications

SWIFT-3D has been applied to several differentproblems in network visualization. These include theability to provide an abstraction that permits visual-ization of the data across the information strata ofnetwork element, network, services and customers;the ability to view cross network interactions andtheir impact upon a service and/or customer; the ca-pabilities to discern impact on one or more customerswhen there is a network event.

An interesting example is the examination of callsthat cannot be completed due to congestion at the

5

customer premise. Keeping this number low is im-portant due to the resources consumed. It is impor-tant both to the customers (who need reliable servicefor telemarketing sales and customer support) andto AT&T from a financial standpoint (because unan-swered calls consume network resources and may in-cur cross-carrier settlement charges without creatingrevenue). In visually exploring voice network events,we noticed that on several days within an interval ofseveral weeks, many unanswered calls originated ina certain metropolitan area (cf. Figure 4). The eventsalways occurred at bottom of the hour (:30) for sev-eral hours in the evening. By interactive queryingwe found that most of the calls were directed at one800 number, and that the number belonged to a radiostation. By tuning in, we discovered that the stationwas giving out free tickets for an upcoming concert.The winner was the tenth caller at the bottom of eachhour.

Another application concerns analysis of an Inter-net service. There is considerable motivation forunderstanding relationships between usage of mul-tiple services, both from a single service provider,and between competitors. AT&T wanted to knowhow much coverage an Internet access service had.The coverage is measured by the number of the areacode and exchanges (and ultimately households orcustomers) where connecting to the Internet is a lo-cal call (usually without per-minute charges). Wecontracted two companies that claimed to have suchdata. We gave them the locations of the modempools and asked them to tell us what codes and ex-changes were covered. We received two very differ-ent answers. To understand the differences we usedSWIFT-3D: areas claimed to be covered in the an-swer of company A were colored blue, those claimedto be covered in the answer of Company B were col-ored green, where both companies agreed, the mapwas colored gray. We noticed widespread differ-ences in many states, while a few states had goodmatches. In order to decide which company’s answerwas more correct, we superimposed our customer us-age data on the map. In the generated visualizations(cf. Figure 5), we saw that there was a lot of us-age in gray and blue areas, but very little usage ingreen (and almost none in black areas). Our conclu-sion was that the answer by company A was more

correct. It further became clear that individual cus-tomers are very aware of local calling areas, and arenot willing to use an ISP when theaccess would betoo expensive. A side effect of our findings was thatthe business decided to not even advertise this ser-vice in areas not covered by local access.

A third application involves recognizing the charac-teristics of virtual private networks (VPNs) provi-sioned by customers on a large packet network, andtheir relationships to physical network facilities. Fig-ure 6 shows the peak volume of Permanent VirtualCircuit (PVC) traffic, by VPN, for the whole networkin one 5-minute period. The display highlights thePVCs having the greatest load. The eventual goalof this study is to understand customer-focused, nearreal-time management of packet networks.

Figure 4: Network event analysis

It should be noted that, even with the disparityamongthese application domains, it was not difficult to tai-lor SWIFT-3D to each. In large measure, this is dueto the similarity of the visual models, and to the high-level descriptions used to specify much of the analy-sis and display. Probably the most difficult aspect inmodifying SWIFT-3D for an application is construc-tion of tools to massage the application’s data into aformat suitable for SWIFT.

6

Figure 5: Geographic coverage analysis

Figure 6: Packet traffic analysis

5 Graph Surfaces

As the size of the data sets increases, the optimiza-tions within SWIFT-3D and the use of large displaysbecome inadequate. What is needed is a new wayto visualize aggregated data. One approach is touse graph surfaces[1], which provide a new visualmetaphor for large networks. They provide a col-lection of natural operations for browsing at differ-ent levels of detail, and fit well in the context oflarge networks, allowing convenient incremental up-date and navigation of external memory graphs (cf.[2]) whose vertex sets are hierarchically labeled, asis typical with call detail and Internet data.

The main idea is to view a network as a discretiza-tion of a 2D surface in 3-space. For a fixed orderingof the network components, a rectangular domain istriangulated and each point is lifted to a height cor-responding to aggregated attribute of the underlyingnetwork. This provides a piecewise linear continuousfunction forming a polyhedral terrain. The terrain isused as an approximation to a surface representingthe communication network. An example is shownin Figure 7.

Figure 7: A graph surface

In order to handle very large communication net-works, a hierarchy of surfaces is constructed. Eachrepresents an aggregate view of the network at alower level. Operations are provided that allow theuser to navigate the surfaces.

7

5.1 Constructing and navigating a graphsurface

A network, during some interval of time, can beviewed as a collection of vertices, a collection ofedges connecting the vertices, and weights on theedges, measuring some attribute of an edge. For ex-ample, in the case of call detail data, the vertices cor-respond to phone numbers.

An important aspect of these networks is that the ver-tices can be viewed as having an underlying hierar-chical structure. This induces a hierarchical structureon the edges. At the lowest level, one has the orig-inal network. Subsequent levels are obtained by co-alescing disjoint sets of vertices at a previous leveland aggregating their corresponding weighted edges.Each level in the hierarchy can be represented as asurface and the hierarchy can be used as a road mapto move from surface to surface. By collapsing sur-faces into edges and expanding edges into surfaces,one gets zoom in and zoom out operations, as ex-hibited in Figure 8. An important point, for largenetworks, is that the hierarchy can be constructed ef-ficiently. (See [1] for details.)

Figure 8: Graph surface navigation

5.2 Implementation

To handle large networks effectively, there are threecases to consider:

� The hierarchy fits in main memory

� The hierarchy does not fit but the vertex set does

� The vertex set does not fit

In the first case, the edges are read in blocks and eachone is filtered up, through the levels of the hierarchy,until it lands in its final level. This can be achievedwith one pass over the data. The second and thirdcase are handled by setting up a multilevel externalmemory index to represent the hierarchy. Filteringthe edges requires now several passes over the data.The I/O performance depends strictly on the I/O ef-ficiency of the access structure.

The visualization of a data set that is several orders ofmagnitude more than the available screen resolutioncalls for some form of hierarchical graph decompo-sition. Graph surfaces by definition provide a geo-metric view of a very large graph in a manner thatis amenable to hierarchical browsing via the zoom inand zoom out operations. This makes it feasible toexplore deeper hierarchy elements at a finer level ofdetail. Also, a graph surface provides both a uniformlocal and global view.

The graphical engine generates polyhedral terrainsthat correspond to individual edges in hierarchy. Ini-tially, a suitable level in the hierarchy is chosen asthe root, depending of the available screen resolu-tion. The polyhedral terrains are generated with a tri-angulation algorithm and are displayed using severalvisual cues and dynamically generated labels. Thegraphical engine is implemented in C++ and uses theOpenGL standard library for the rendering portion.Currently, a mouse/keyboard interface is used. Theuse of joysticks and gestures to navigate the environ-ment is currently under consideration.

5.3 Applications

Currently, graph surfaces are being used experimen-tally for the analysis of several large multi-digraphs

8

arising from the AT&T network. These graphsare collected incrementally. For example, in thecall-detail multi-graph, vertices correspond to phonenumbers and edges represent a call from one numberto another. The graph grows by daily increments ofabout 300 million edges, defined on a vertex set onthe order of 300 million vertices. The aim is to pro-cess and visualize this type of multi-digraph at a rateof at least a million “edges” per second.

Internet data is another prime example of a hierar-chically labeled multi-digraph that fits the graph sur-faces metaphor quite naturally. Eachi-slice repre-sents traffic among the aggregate elements that lie atthe i

th level of the hierarchy. The navigation opera-tions can be enhanced to perform a variety of statis-tical computations in an incremental manner. Thesein turn can be used to animate the traffic behaviorthrough time.

When the vertices of the multi-digraph have an un-derlying geographical location, they can be mappedinto a linear order using, for example, Peano-Hilbertcurves, in order to maintain some degree of geo-graphical proximity. In this way, the constructed sur-face maintains a certain degree of correlation withthe underlying geography.

6 Conclusions

The Infolab work, with its heterogeneous displays,its flexible and powerful software base, and its newvisual metaphors, has proved successful in giving itsusers a new way of viewing AT&T networks, andserves as a starting point for much future work.

At the simplest level, the system needs to be pack-aged so that it can be used on a daily basis by net-work analysts and administrators. The system shouldbe configured so that a user can easily modify it forspecialized display and analysis needs. This work isalready underway for applications in NOC-2000 andthe AT&T frame relay network. In addition, the sys-tem needs to be available on the desktop. To addressthis last point, we have a prototype implementationin Java and Java-3d, where the user interfaces runson a PC and more powerful servers handle data pro-cessing.

Another avenue of work is to enhance the system tobetter handle the heterogeneity of AT&T’s networksand workplace. It should be possible for employees,at different locations with different display environ-ments, to visually share network information. At thesame time, the many types of networks could be log-ically integrated into a single network.

Our display wall has proven itself. However, its cur-rent set of input devices (mouse and keyboard) seemsinadequate. We would like to explore new sources ofinput, such as wands, gesture recognition and palmcomputers.

Many of the views in the system are, at present, basedon the underlying geography. This is a significantlimitation for network visualization. Often, manyendpoints are located in a few dense metropolitanareas, with large ‘deserts’ in between, so maps donot use the available pixels efficiently. Also, mapswork well for displays of endpoints (vertices) butnot so well for edges (vertex pairs), let alone groupsof more complex structures such as routes, flows,or subgraphs representing virtual networks. WithIP networks, geographic coordinates are not alwayseven available. One can move to more abstract topo-logical representations, but the standard visualiza-tion techniques rapidly become unusable as the datagrows, even though a megapixel display delays thissomewhat. There is need for additional work in dis-play metaphors and interaction techniques for largegraphs. Graph surfaces, described above, area oneapproach, but additional new ways of “looking” atnetworks and network data are called for.

7 Acknowledgements

For collaboration on experiments and support oftechnology transfer, the authors thank Ed Berberich,Chee Ching and John Ennis from NOC-2000, andArvind Chakravarti, Ram Chelluri, Quynh Nguyenand John Medamana from the AT&T Frame RelayNetwork lab.

9

References

[1] J. Abello and S. Krishnan. Graph Surfaces. InInt. Conf. Industrial and Applied Math., July1999.

[2] J. Abello and J. Vitter, editors.External Mem-ory Algorithms. AMS-DIMACS, 1999.

[3] C. Ahlberg and E. Wistrand. IVEE: An En-vironment for Automatic Creation of DynamicQueries Applications. InProc. ACM CHI Conf.on Human Factors in Computing (CHI95),1995.

[4] C. Ahlberg and E. Wistrand. IVEE: An Infor-mation Visualization and Exploration Environ-ment. InProc. IEEE Int. Symp. on InformationVisualization, pages 66–73, 1995.

[5] Richard A. Becker, Stephen G. Eick, and Al-lan R. Wilks. Visualizing network data.IEEETransactions on Visualization and ComputerGraphics, 1(1):16–28, March 1995.

[6] A. Buja, J.A. McDonald, J. Michalak, andW. Stuetzle. Interactive Data Visualization Us-ing Focusing and Linking. InProc. IEEE Visu-alization, pages 156–163, 1991.

[7] C. Cruz-Neira, D. Sandin, and T. DeFanti.Surround-Screen Projection-Based Virtual Re-ality: The Design and Implementation of theCAVE. In Siggraph 1993 Conference Proceed-ings, pages 35–142, 1993.

[8] Paul Woodward et al. University of MinnesotaPowerWall, 1998. http://www.lcse.umn.edu-/research/powerwall/powerwall.html.

[9] Daniel A. Keim and Annemarie Herrmann. TheGridfit Algorithm: An Efficient and EffectiveAlgorithm for Visualizing Large Amounts ofSpatial Data. InProc. IEEE Visualization,pages 181–188, 1998.

[10] Eleftherios E. Koutsofios, Stephen C. North,and Daniel A. Keim. Visual exploration of largetelecommunication data sets.IEEE ComputerGraphics and Applications, 19(3):16–19, May1999.

[11] S. F. Roth, P. Lucas, and C. C. Gomberg. Vis-age: Dynamic Information Exploration. InProc. ACM CHI Conf. on Human Factors inComputing (CHI96), April 1996.

[12] M. O. Ward. XmdvTool: Integrating Multi-ple Methods for Visualizing Multivariate Data.In Proc. IEEE Visualization, pages 326–336,1994.

10