Volume 8, Number 2 Department of Computer Science Fall, 1999 Brown University

conduit!condu t!

Eli Upfal

RANDOMNESS, PROBABILITYand COMPUTING

om the highly theo-

“God does not play dice with theuniverse” was Albert Einstein’sfamous objection to modernquantum-mechanics theory. Thisand similar objections reflect thehuman conceptual difficulty ingrasping a world that is not gov-erned by fully deterministiclaws. Nevertheless, research inthe past century has demon-strated the dominance of randombehavior and statistical laws inalmost any field of science,ranging from sub-particle phys-ics to free market economics.Computer science is no excep-

“how can a string of randomnumbers (perfect noise) that

adds no information to the pro-gram actually improve the per-

formance of its execution?”

tion: frretical notion of probabilistictheorem proving to the verypractical design of PC Ether-net cards, randomness andprobabilistic methods play adominant role.

Roughly speaking, random-ness comes up in two aspectsof the study of algorithms: ra

ndomizedalgorithms and probabilistic analysis ofalgorithms. Randomized algorithms arealgorithms that make random choicesduring their execution. In practice, arandomized program uses values gener-ated by a random number generator todecide the next step at several branchesof its execution. For example, theEthernet access protocol, implemented

Brown University, Box 1910, Pr

in the Ethernet card, uses random num-bers to decide the exact timing by whichit tries to access the Ethernetcommuni-cation medium. Other commonly usedapplications of randomized algorithmsare Monte-Carlo simulations and pri-mality testing for crypto-graphy applica-tions. In these and many otherimportant applications, randomized al-gorithms are significantly more efficientthan the best known deter-ministic solu-tions. Furthermore, in most cases therandomized algorithms are also simplerand easier to program.

A randomized program can be viewed asa conventional (deterministic) programthat reads two streams of inputs: one isthe actual input of the program, theother is a sequence of randomly gener-ated numbers. This view raises an in-

triguing philosophical question: how cana string of random numbers (perfectnoise) that adds no information to theprogram actually improve the perfor-mance of its execution? Theoretical com-puter science still does not have a fullexplanation of that phenomenon, thoughmore and more applications take advan-tage of it.

ovidence, RI 02912, USA

Another algorithm-related application ofprobability theory is in analyzing the av-erage-case performance of algorithms.Algorithm and complexity theory tries toclassify computation problems ac-cording to their computational com-plexity, in particular distinguishingbetween easy and hard to solve prob-lems. For example, complexity theoryshows that the famous travelingsalesman problem is NP-hard. Thus it isvery unlikely that it has a solution sig-nificantly faster than enumerating allpossible routes—a task that is exponen-tial in the number of cities and thus notfeasible for a large number of cities.

An embarrassing phenomenon for theclassical ‘worst-case’ complexity theoryis that problems classified as hard tocompute by the theory are often easy tosolve in practice. Probabilistic analysisgives a theoretical explanation of thatphenomenon, namely that these prob-lems are hard to solve on some small setof pathological hard inputs but are actu-ally easy to solve on most inputs, partic-ularly those that come up in real-lifeapplications—scheduling, the travelingsalesman problem, packing and coveringare just a few examples.

Dynamic AlgorithmsThe focus of my research is the applica-tion of probability theory in computerscience. In particular, I am currently in-

“Worst-case analysis rarelygives interesting insight into

the actual performance of adynamic algorithm”

terested in studying dynamicalgorithms through sto-chastic an-‘alysis. Researchin theoretical computer sci-ence has focused mainly onstatic computation problems,where the input is known be-fore the start of the computa-tion and the goal is to minim

ize thenumber of steps till termination with acorrect output. However, many impor-tant processes in today’s computing aredynamic or interactive processes, where-by input is continuously injected to thesystem and the algorithm is measuredby its long-term steady-state perfor-mance. Examples of dynamic processesinclude communication protocols, mem-ory-management tools, and time-sharingpolicies. The goal is to develop new toolsfor designing and analyzing the perfor-

conduit

mance of dynamic processes, in partic-ular through modeling the dynamicprocess as an infinite stochastic process.

A key feature of dynamic algorithms isthe need to make decisions ‘on-line’without information about future re-quests. An obvious example is cachemaintenance protocols. A cache is a fastmemory that serves as a buffer betweenthe processor and a slower but largermemory module. Since a cache is typi-cally small, the program often requestspages that are not in the cache. When anew page is brought to the cache, someother page must be returned to the mainmemory. Accessing that page later willbe expensive, and thus the algorithmtries to evict from the cache pages thatare less likely to be requested in the fu-ture. The algorithm makes these deci-sions, of course, without actually know-ing the future sequence of requests.

As in the case of static algorithms, ran-domness is introduced into dynamiccomputation through the algorithm, theinput or both. Many interesting dynamicprotocols, such as the Ethernet accessprotocol mentioned before, are random.An execution of a dynamic random algo-rithm, even on a fixed input sequence,defines an infinite stochastic process inwhich a state at a given step depends onthe history of the process. Analysis ofsuch a process requires a different ap-

proach and different tools than thoseused in analyzing the finite execution ofa randomized static computation.

Stochastic Analysis of DynamicAlgorithmsWorst-case analysis rarely gives inter-esting insight into the actual perfor-mance of a dynamic algorithm. A worst-case adversary can generate extremelyhard sequences of requests, and the per-formance of the algorithm on thesepathological cases does not accurately

! 2

Gopal Pandurangan

represent its efficiency. To offset the ef-fect of rare cases, it is useful to analyzethe performance of dynamic systemsunder some stochastic assumptions onthe stream of inputs. Such assumptionsare more realistic in dynamic settings,especially when requests are originatedby a number of independent processors,than in static analysis. The stochasticprocess that controls the stream of re-quests might be stationary, periodic, oreven bursty. The goal is to obtain resultsthat are valid under the weakest set ofassumptions. The advantages and prac-ticality of this approach have been welldemonstrated by the achievements ofqueuing theory. Our goal is to apply sim-ilar techniques to dynamic computerprocesses that do not fit the queuingtheory settings.

Stochastic analysis of dynamic processesbuilds on the rich theory of stochasticprocesses, in particular queuing theory,and the theory of stationary processes.However, in many cases new tools areneeded to address the specific problemsposed by computer-related processesthat are discrete and involve compli-cated dependency conditions.

Recent WorkPhD candidate Gopal Pandurangan andI are studying an improved protocol foradmission control in fast communicationnetworks. Modern communication proto-cols such as ATM (asynchronoustransfer mode) achieve high utilizationof channel bandwidth by multiplexingcommunication streams with differentflow characteristics into one communica-tion channel. Requests for communica-tion are submitted to the networkmanagement protocol with some statis-tical characterization of the requiredcommunication. The network (flow)management protocol uses this informa-tion to statistically multiplex as manycommunication requests as possiblewhile maintaining global network per-formance. Next-generation communica-tion networks are expected to provideQoS (quality of service) guaranteeswhen satisfying communication re-quests. In particular, QoS protocol is ex-pected to limit to a pre-specified valuethe probability of communication failure

conduit!

due to events such as link and bufferoverflow, packet loss, jitter, etc.

Our new protocol employs an efficientMonte-Carlo method for estimating thefailure probability of a network. Thetight estimate allows the protocol to uti-lize the network bandwidth fully with-out violating the QoS requirements. Thenew method is particularly useful in adynamic setting in which communica-tion requests are dynamically added toand eliminated from the system. Theamortized cost in our solution of up-dating the estimate after each change isproportional to the fraction of links in-volved in the change rather than to thetotal number of links in the network. Weare currently running an extensive sim-ulation project to evaluate the perfor-mance of the new technique undervarious network and load settings.

CourseThe research area combines two disci-plines, algorithm theory and probabilitytheory. My new course ‘CS155: Probabi-listic Methods in Computer Science’ ex-poses students to the interplay betweenthese two areas. They study basic proba-bility theory, in particular discrete prob-ability theory (which is more relevant tocomputer science applications) and thenapply this theory to design and analysisof randomized computer algorithms for avariety of applications.

3

INSIDE IJCAI ’99

Tom Dean in Stockholm

Back, l to r: David Carlson, Roberto Tamassia, Eli Upfal. Front, l to r:Charles Fiduccia, John Savage, Franco Preparata

SAVAGE HONORED ON HIS 60th

Nearly four years ago, atthe 1995 IJCAI (Interna-tional Joint Conference onArtificial Intelligence) con-ference in Montreal, I wasasked to be program chairfor the 1999 IJCAI inStockholm and also toserve on the board oftrustees for IJCAI Inc., thegoverning board that runsthe IJCAI conferences,oversees a trust fund, de-termines conference ven-ues, awards prizes, andgenerally plays an impor-tant role in the fieldthrough the various activi-ties associated with the bi-ennial IJCAI conferences.

conduit!

IJCAI is the premier international con-ference in AI and the invitation fromthen-president Ray Perrault was veryflattering. Stockholm was one of my fa-vorite cities, I enjoyed working with Rayand the other trustees, and I thoughtthat my experience running other con-ferences (in addition to smaller confer-ences and workshops, I was programchair for NCAI in 1991—the NationalConference on AI, often called AAAI forthe North American professional associ-ation that runs it) would make this newtask relatively straightforward. I naivelyagreed to serve.

I can’t begin to say how unprepared Iwas for the experience. NCAI is man-aged by a well-run professional organiza-tion with relatively deep pockets. Muchof what was necessary in running AAAI-91 was handled without my being awareof it by Carol Hamilton, the ExecutiveDirector for AAAI, and her excellent

On September 21, in honor of John Savage’s 60th birthday, Franco Preparata hosted aone-day technical forum in the department on ‘‘Algorithmic Research.’’ The event recog-nized, in its subject and character, John’s constant devotion to productive and well-regarded academic research. Colleagues and former doctoral students dedicated to thehonoree some of their more recent work. The department participated actively in theevent, recognizing John’s many contributions as a co-founder of our department and hisassiduous stewardship as chairman and excellent citizen. After initial remarks by Tom

Dean and Andy van Dam, Franco

4

kicked off the day with a talk on“Reconstructing a sequence from itssamples: DNA sequencing at theinformation-theory bound,” CharlesFiduccia (PhD ’73, Division of Engi-neering) spoke on “Optimal mono-tonic search,” Eli Upfal on“Reducing network congestionthrough balanced allocation,” DavidCarlson (PhD ’80) on “New insightsinto two old algorithms: the FastFourier Transform and metropolisMonte-Carlo method,” and RobertoTamassia on “Graph drawing andinformation visualization.” Theintense technical program had arelaxing break during the buffetluncheon and was capped by a gra-cious departmental reception (withbirthday cake, of course).

“Runconfe

all sorto mapp

a

staff. I was oblivious of many of the de-tails of running a large conference. Eventhough IJCAI is a larger conference thanNCAI and much more complicated in itsinternational focus, most IJCAI work isdone by the trustees and numerous vol-unteers from the academic and researchcommunities. The trustees turn over reg-ularly, so corporate memory is short. Forthe most part, each IJCAI is inventedanew, with little help from past experi-ence.

Each IJCAI has a program chair incharge of the technical program, a gen-eral chair in charge of the whole process,a conference arrangements chair, and nu-merous other local and special programchairs. I worked most closely with Prof.Luigia Carlucci (Gigina) Aiello, the gen-eral chair and a professor at Universitàdi Roma “La Sapienza,” and Prof. AnitaKollerbaur, the conference arrangementschair and a professor at Stockholm Uni-versity and Royal Institute of Technology.Gigina and Anita were wonderful to workwith, but our job was incredibly frus-trating and a bit unnerving as we tried tofigure out how to run a conference with amore than million-dollar budget (and nosmall amount of financial risk) startingpretty much from first principles.

ning an internationalrence is fraught with

ts of politics. You haveake sure that all theropriate nationalitiesre represented”

Well, we didn’t really start from firstprinciples; Anita knew a lot about plan-ning large projects (part of her academicexpertise) and she drew upon the skills ofthe administrative and technical staff ather university and hired a professionalconference organizer to help with variousaspects of the planning and execution. Iarranged for Carol Hamilton and AAAI tohelp with the extensive correspondenceand paper-handling associated with thetechnical program. In fact, we had to puttogether a small army of people to help

conduit!

with the conference, and Gigina, Anitaand I had to direct the troops.

Running an international conference isfraught with all sorts of politics. Youhave to make sure that all the appro-priate nationalities and professional so-cieties are represented. Even though theconference was to be held in Stockholm,it was supposed to be hosted by all theScandinavian countries. Registrationtypically doesn’t cover the cost of run-ning a conference, and so to break evenwe needed to get sponsors willing to helpout financially but not require their cor-porate logos to be tattooed on the fore-heads of the attendees.

Much of my effort was directed at put-ting together the program committee,finding chairs for workshops and tuto-rials, and choosing invited speakers andcajoling them to attend even though wecould only cover a portion of their overallexpenses. I enlisted the help of 37 top re-searchers to serve on the program com-mittee, and these folks in turn enlistedthe help of nearly 400 reviewers tohandle the almost 800 submissions andover two thousand reviews. I won’t boreyou with the details of coordinating thiseffort and coping with inevitable glitchesand embarrassments that occur whendealing with so many people and tryingto adhere to a strict deadline determinedby printers, international shipping rules,and of course the actual date of the con-ference at which the final proceedingsare to be handed out to attendees.

By the end of last February, we had se-lected 195 papers for presentation, 14 in-vited speakers were lined up, 29 work-shops and 20 tutorials were on the pro-gram, and we had made the selectionsfor the Computers & Thought and Re-search Excellence awards. The locationsfor the conference, the receptions, thebanquet and other events were settled.Given the chaos of reinventing IJCAIand the fact that none of us were exactlysure who was responsible for what, I wasrelieved that Gigina, Anita and I werestill speaking with one another. Afterwhat seemed like two years of steadywork, I thought my job was over.

Unfortunately, there were some little de-tails yet to go: putting together the pro-ceedings, developing the content for the

5

brochure, providing a schedule for theconference program, creating paper ses-sions and assigning them to rooms,finding reliable session chairs and peopleto introduce invited speakers (92 people),and a myriad of other details. It seemedthat I would never get out from underthe yoke of IJCAI.

My wife Jo and I arrived in Stockholm onJuly 30, several days before the officialbeginning of the conference. There weretutorials and workshops to check on, arobotic soccer competition that was co-lo-cated with IJCAI, and trustees’ meetingsto attend. Indeed, trustees’ meetings con-tinued throughout the conference, deal-ing with the previous conference (inNagoya), the current conference, thenext conference (which is to be held inSeattle and for which the planning was

Skeppsholmen

well under way), deciding on the locationfor 2003 (Mexico), and dealing with var-ious aspects of planning for 2005, 2007and beyond.

The weather was perfect. Stockholm is abeautiful city and it is great fun to tourits shops and museums, flit from islandto island on its convenient water taxis,take day trips out to the archipelago, andrelax in its ample parks and gardens.With the exception of a few rain showers,the perfect weather, dry and warm butnot hot, continued throughout. The con-ference attendees were treated to a boattrip and a banquet on a fortress island.It stayed light till 10 pm or so and most

conduit

people took advantage of the light towander in the old town (Gamla Stan),enjoy a walk around Skeppsholmen, ortake in the rides and other amusementsin Djurgarden. The trustees plotted andplanned and were rewarded for their ef-forts by a wonderful dinner in a restau-rant located in one of the royal gameparks. There were more courses than Icould count and fabulous wines that I’llnot likely sample again any time soon.One of the trustees knew a little aboutwine and the steward took us to visittheir cellar; along one wall was a collec-tion that included one bottle from each ofthe major houses in the Bordeaux region,starting from 1900 and extending to thepresent. The only missing years werethose in which war had eliminated anentire year’s production.

I heard very few technical papers atIJCAI ’99 and, though I heard parts of al-most every invited talk, for one reason oranother I was never able to sit throughan entire talk. After the last of the tech-nical sessions, I felt completely drained.Luckily several of Jo’s family had arrivedin Stockholm the previous day and so Ihad plenty to distract me from my post-partum blues. Jo’s sister Nancy and herdaughter Lila had flown in from Paris,Nancy’s husband had flown in fromBoston, and Lila’s boyfriend came fromRome right after he finished work onFriday. We spent a pleasant weekend en-joying what the conference’s attendeeshad been experiencing the week before.

In retrospect, it was a grueling fouryears. The final result was very grati-fying but I don’t know that I would do itagain. Delayed gratification is commonin the academic world: you submit apaper or proposal and wait months for aresponse. At this point in my life, how-ever, I think I want my feedback a littlemore immediately. Running a confer-ence like IJCAI is a bit like building ahouse: you really can’t enjoy the houseuntil the roof is on, the plumbing work-ing, the walls painted, and it’s ready tomove in; in the interim, there is alwayssomething to do, most of it tedious. Still,I have to admit that it’s pretty excitingwhen it all comes together. Todaysomeone called me about running an in-ternational conference on ...

! 6

PETER WEGNER ON THE MEND

Peter and Judith Wegner at the Rehabilitation Hospital of RI in N.Smithfield, where Peter had been undergoing physical therapy.

Happily, he is now back at home on the East Side

The following article and photograph above areby senior news writer Kristin Cole. Theyappeared in the October 15-21 issue of theGeorge Street Journal.

The music of Brahms filled the Londonhospital room where Peter Wegner, a re-tired Brown computer science professor,had only recently regained consciousnessafter being in a coma for four weeks.

The bus that struck him as he walked to-ward Trafalgar Square a month earlierhad smashed an elbow, broken ribs, andresulted in a head injury from which doc-tors had initially predicted only a 5- to10-percent chance of survival.

When asked whether he liked Brahms,Wegner replied, “Yes, especially BostonBrahms” —a pun referring to the upper-class intellectual Brahmins of 19th-cen-tury Boston. For friends and colleagueswho had given Wegner a chair engraved“to an unrepentant punster” as a retire-ment gift, it was perhaps the surest signthat he was on the road to recovery. “Itwas reassuring to know it was the sameold Peter,” said Judith Wegner, his wife of43 years. “He beat the odds.”

conduit

Wegner is now at the Rehabilitation Hos-pital of Rhode Island in North Smith-field. He returned to the United StatesSept. 30, nearly four months after the ac-cident. On June 11, Wegner had beenplanning to take a walk in TrafalgarSquare, one of his favorite haunts. Hehad just retired after 30 years of teachingat Brown and was looking forward to asummer filled with engagements.

His first stop in England was to attend areunion of Jews who, as children, had es-caped Nazi terror in their homelandswhen they were accepted by England aspolitical refugees. Wegner had left Vi-enna as a 6-year-old, one of 10,000 chil-dren on the “Kindertransport.” After thereunion, he was slated to deliverspeeches to professional conferences inPortugal and Scotland. To cap the Euro-pean trip, Wegner was to receive Aus-tria’s highest academic award, for hislifetime contribution to the field of com-puter science.

His wife had left Wegner only a shorttime before he headed to TrafalgarSquare, returning to her sister’s home,where they were staying, to help preparedinner. But when the meal was readyand Wegner had not arrived home, thefamily began to worry. They notified po-lice and learned a short while later thatWegner had been taken by helicopter

! 7

Said Ed Lazowska in his And

(Not ZZ Top) Andy andSteve Carmody, Wolfg

Scenes from las

from the accident to the Royal LondonHospital. His wallet had not contained in-formation for contacting his sister-in-law.But the wallet did contain names of col-leagues, including Brown faculty mem-bers Andries van Dam and ThomasDoeppner.

Initial reports about Wegner’s conditionwere pessimistic, and the couple’s sons,Mark, Jeremy and Michael, flew from theUnited States to London to be at the hos-pital. In the weeks that followed, thefamily updated friends and colleagues onWegner’s status by e-mail, maintaining alist that reached across the Atlantic, in-cluding many at Brown. Additionally, Ju-dith notified her husband’s professionalcommitments of the situation.

In the hospital room, Judith and hersister, Marion Rosenberg, sang familiarfolk songs to Wegner in English andGerman, in which he is fluent. They cele-brated his 67th birthday there, paperingthe walls with cards that wished him welland often contained puns. When Wegnerfirst came out of the coma, a tube in histhroat prevented him from speaking, butit was obvious he could understand whatwas being said and recognized his family,said Judith. “That is an unusually longcoma for someone to recover from. He is afighter. He is a determined person and heis working very hard at it now.”

conduit

yfest kick-off speech, “It’s hard to believe Andy is 60.

, l to r, Hendrik-Jan Thomassen,ang Millbrandt and Ken Sloan

Gue

t May’s “Andyfest,” a two-day s

In therapy Wegner has strengthened hisability to walk, read and write. A recentmilestone in his recovery was walkingaround the perimeter of the hospitalbuilding without the aid of a cane. “Ju-dith has helped me enormously duringthis time,” he said. Wegner looks forwardto the day he will return to his office; al-though retired, he still plans to continuehis research. There are also retirementdreams to satisfy, said Judith, such ascruises to Alaska and the Greek Islands.And he has yet to receive the AustrianMedal of Honor for Science and Art,which the government offered to confer inthe hospital. Wegner refused. He wantsto travel to the country to receive theaward, as planned.

They are all plans that friends and col-leagues thought Wegner would somedaybe able to satisfy after learning that hisability to make puns was intact only daysafter coming out of the coma. “It certainlyshowed the sign that he was recovering,”said Doeppner, associate professor ofcomputer science. “He can come up withthe most amazing puns in zero time.” Ju-dith maintains a list of Wegner’s recentpuns at their home in Providence, wherethe chair with the engraving “to an unre-pentant punster, indefatigable scholar,and generous friend” awaits his return.

! 8

He looks only 45, same as he did when he was 30!”

sts seated for lunch on campus

ymposium in celebration of

VISUALIZING PROGRAM EXECUTION

Steven Reiss

Enjoying an outdoor bMillan building where

Andy’s 60th birt

Introduction

Today’s software systems are complex,and their behavior, i.e. what happens asthey execute, is typically even more so.Here one must deal with the interactionof large amounts of code, distributed com-putations, external systems, multiplethreads of control, locks, asynchronousevents, message traffic, and other compli-cating issues. However complex this be-havior may be, programmers must beable to understand what is going on, espe-cially when the systems start to act in anunexpected or erroneous manner.

Our approach to this problem of softwareunderstanding is to use information gath-ered as the system executes to give theprogrammer sophisticated visualizationsdescribing the execution. The amount ofdata that can be collected is vast and caneasily be overwhelming. Visualizationuses the brain’s sophisticated recognitionabilities to quickly find relevant patternsin a sea of data in order to make execu-tion understanding practical.

Our research here involves four relatedbut separable problems. The first involvesobtaining and storing the data. Thesecond involves letting the programmerspecify what needs to be visualized to un-

conduit

reak in front of the new Mac- most of the talks took place

President Gof the Unive

hday, entitled “The Computer, the

derstand a particular aspect of the be-havior of a particular system. The thirdinvolves processing the raw trace data soas to obtain the information needed forthe visualization. The final problem in-volves actually producing a visualizationwith which the programmer can interactto obtain the desired insights.

Obtaining the Data

A lot of data is needed to describe the ex-ecution of a system. This includes func-tion calls and returns; thread creation,destruction, and state changes; the stateof locks and other synchronization mech-anisms; memory access and paging be-havior; local control flow; timing infor-mation, both real-time and execution-time; message passing; files loaded andused; and file and socket utilization. Thegeneral rule for software understandingis that you want to collect as much dataas possible since there is some potentialproblem for which that data will be thekey to understanding.

We currently have two separate systemsthat can gather much of this trace data.The first, AARD, works with C or C++programs. It consists of a package Wolfthat takes executable files and rewritesthem with instrumentation code, and a

! 9

ee presenting a handsome photographrsity at the beginning of the banquet at

the Biltmore

Academy, and the World.”

Over 300 partici

One of Andy’s first doctolands a good one a

package Vark that uses the raw instru-mentation data to provide data for visual-ization. The second system, TMON, usesJava’s JVMPI interface to collect trace in-formation on Java programs. It handlesmultiple threads of control withoutadding any additional run time synchroni-zation and records everything that JVMPIoffers.

Both of these systems generate largeamounts of data. Although it will varywith different applications, our currentbest guess is that they typically produceabout 1G of raw trace data for every 10seconds of raw CPU time. (This is for rela-tively high-level information such asmethod calls; with low-level informationsuch as basic block entries or memory ac-cesses, this goes up by at least an order ofmagnitude.) Our research here involvesmanaging this trace data so that we candeal with large, long-running systems.

We take two different approaches here.The first involves using the increase incomputer capabilities to provide a hard-ware-based solution. We are looking atdoing traces on machines that have sev-eral gigabytes of memory, multiple proces-sors, and high-speed connections to largeamounts of RAID-based disk storage. Wehope that the combination of such mas-sive workstations with sophisticated tracecollection code will let us collect completetraces with minimal impact on the run-ning program.

conduit

With Ed Lo

pants attended Andyfest at venu

ral candidates, Ingrid Carlbom,t the roast after the banquet

The second approach involves under-standing what information is actuallyneeded for a particular visualization anddoing the appropriate analysis and pro-cessing as the raw data is generated sothat only the processed data is stored.One difficulty here is that we want to letthe programmer define what should be vi-sualized and hence have to compute thenecessary filters and processors dynami-cally. A second difficulty is that the pro-cessing can be quite sophisticated, in-volving tracking execution histories,maintaining the execution state, andlooking at the operations of multiplethreads without adding additional syn-chronization that would change how theprogram is behaving.

Determining What to Visualize

One conclusion we drew from our pastwork on software visualization was thatno single or small set of fixed visualiza-tions will address the wide range of ques-tions asked as a programmer attempts tounderstand the behavior of a complexsystem. Rather than attempting to definesuch a fixed set and having it not be par-ticularly useful, we have focused on let-ting the programmer define what shouldbe visualized and how it should be dis-played. By making the foundation of sucha system rich enough, we hope to addressa wide range of understanding problemsand thus make our system practical.

! 10

azowska ’72 and John Hughes, whorchestrated the entire event

es both on- and off-campus

“Raw traceuseful. No

much inforingful disprelevant to

typi

To this end, we created, as part of ourDesert environment, the CACTI front endfor defining visualizations. CACTI lets theprogrammer define the data to visualizeby graphically defining sets of objectsusing fields that are the various domainscomposing the raw data. It is, in effect, auniversal-relation-assumption-based vis-sual query language that can access bothdatabases describing the static structureand symbols of the system and the var-ious dynamic analyses that are availablefrom the processed trace data. It is rela-tively easy to use, as we have demon-strated by addressing a wide range ofspecific understanding problems in-cluding looking for compiler-generatedtemporaries, abstracting the dynamic callgraph, exploring memory behavior overtime, looking for patterns in the call stackover time, and finding methods and vari-ables that are never called.

data by itself is not veryt only is there much toomation to create mean-lays, but the information a particular problem is

cally well hidden”

Our continuing research in this area willattempt to make the CACTI interfacemore intuitive, provide a reasonable frontend to the large number of potential traceanalyses, and handle multiple databaseswith large numbers of potentially over-lapping domains.

Processing the Raw DataRaw trace data by itself is not very useful.Not only is there much too much informa-tion to create meaningful displays (tex-tual or graphical), but the informationrelevant to a particular problem is typi-cally well hidden. To make the datauseful, one must look at the results of oneor more analyses that are run separatelyor concurrently on the same trace beforebeing combined to provide the appro-priate visualization data.

The current analyses we use include pro-filing information (both single-level, à la

conduit!

UNIX prof, and two-level, à la UNIXgprof), raw call data, call dags (con-verting the dynamic call tree into a dagby merging common subtrees), memory-management data, interval analysiswhere the execution is broken into inter-vals in which the data is combined, andinput/output behavior. While all theseare not available with all the currenttrace packages, we have used each ofthem with real trace information to ad-dress specific behavioral issues.

Our experience here is that more andbetter analysis methods are required tomake better use of the trace data. Oneapproach we are taking is to look at anal-yses that deal with specific issues such asmultiple threads of control (e.g. findingpotential deadlocks or race conditions)and message-passing protocols. Anotherapproach is to use techniques developedfor data mining to attempt to find pat-terns, both expected and unexpected, inthe trace data. A third approach involvesproviding a high-level language in whichthe programmer can define the pro-cessing to be done. This last approach isinteresting in that such a specificationcan be used to generate appropriate fil-ters that limit and do some data pro-cessing as the trace is generated.

Viewing the ResultUnderstanding behavior requires notonly flexibility in the information to bedisplayed and the processing to be doneon the raw trace data, but also in a va-riety of different visualization strategies.Different visualization techniques areappropriate to different types of data andemphasize different aspects of the datathat they are presenting. Our approachhere has been to provide a visualizationframework (VALLEY) along with a frontend, MIRAGE, to offer the programmer arange of different styles. The currentsystem provides about ten differentstyles of 3D visualizations including sev-eral file-based visualizations similar toSeeSoft, dot plots, compact trees, graphs,time-based mappings, spiral-based linearviews, and interval analysis.

The CACTI system interfaces with thisframework by letting users select a visu-alization appropriate to the data theyspecified. The system analyzes the data

11

A

and the different visualization strategiesto determine which may be appropriate.The user then can select one of the appro-priate ones and parameterize it, speci-fying properties of the visualization andassociating data fields with visualizationproperties.

To be useful for understanding, however,even the most sophisticated visualizationneeds to be interactive, so that the pro-grammer can browse over the data andthen focus on the most relevant aspects.MIRAGE lets the programmer fly overthe data, change the various parametersdefining the visualization, and, to a lim-ited extent, interact with the visualiza-tion.

Much of our ongoing research in this areainvolves extending these browsing tech-niques to make our visualizations moreuseful. Here we are looking into differentways of filtering the data to be visualizedbefore it is displayed, letting the user dy-namically control what is displayed ornot displayed, and correlating multiplevisualizations of the same data in a syn-

conduit

LUMNI EMAIL

ergistic manner. Another aspect, workbeing done jointly with David Laidlaw, in-volves developing new art-based visual-ization techniques that can convey moreinformation in the limited screen spaceavailable and make patterns in the datamore apparent to the viewer.

ConclusionThis research, a continuation of the soft-ware visualization work done at Brownover the past seventeen years, has beenongoing for the past six years and we ex-pect it to continue for several more beforewe have a truly practical system for un-derstanding complex system behavior.The work is funded in part by the Na-tional Science Foundation, and we are al-ways looking for additional collabora-tions. Here we are interested in identi-fying potential users, establishing re-search partnerships, getting people’sfeedback and experiences with softwarevisualization, and identifying specific be-havioral problems that visualizationmight be able to address.

MARY TAFFS ’75I’d like to announce the publica-tion of my first novel, Martha’sMadness, by Awe-Struck E-Books this September. Mar-tha’s Madness is a love story atheart, but also the story of awoman learning to value andtrust herself. It takes place onthe Oregon Coast and in RhodeIsland. The hero and heroinemet at Brown as undergradu-ates, and the heroine was a com-puter science major. I was verywith the following recent review

pleased

which gave the book four stars and classi-fied it a “Must Read.” Said reviewerAngie Evans of Reviews@SimeGen, “Ienjoyed this delightful book. It was wellwritten and the characters were believ-able. Martha’s Madness is a good read.”A second novel, Celtic Knot, featuring aBrown CS grad as heroine will be pub-lished in a few months. Both books will

be available for sale on diskette as down-loads or as Rocket Editions. For moreinformation, visit http://www.awe-struck.net or my web site at http://www.spiritone.com/~mtaffs/index.html.

I fell in love with the world of booksbefore I could read, and writing is whatI’ve always wanted to do. My computercareer started out at the American Math-ematical Society in downtown Provi-dence. Later, I worked at SofTech inNewport for several years doing Ada com-piler development with my husband Dave(’75). We moved to the Portland, Oregonarea in 1986, and I’ve worked most of thetime since then at Mentor Graphics onDOC, technical publishing softwarewhich was discontinued as a product in1991, but is still supported for the soleuse of Boeing.

I decided in about 1993 that if I was evergoing to write, I’d better get busy NOW.My favorite books at the time were mys-teries and thrillers, so I decided to writean amateur sleuth series. I found out,though, that while I love to read them,I’m not good at coming up with all those

! 12

JO(BS ’70,

John and his wifein Andyfe

plot twists and red herrings, and the rela-tionships between the characters werewhat I most loved writing about. That ledme to move more into the women’s fiction/romance area, and that’s where I trulybelong. mtaffs@SpiritOne.com.

MAARTEN VAN DANTZICH ’93Since leaving Brown I’ve been at Micro-soft Research, exploring the use of 3Dgraphics in user interfaces for desktopapplications. I enjoyed visiting Brown inSeptember and presenting an overview ofour recent work; it’s fun to be the visitingpresenter in Lubrano now, and great tosee how healthy the department is. In fiveyears many of the people have changed,

conduit!

HN GANNON ScM ’72) 1948-1999

Nancy participatingst last May

but it still feels familiar. Not that thereweren’t any surprises: I found a coffee-cart in the lobby of the CIT, which wouldhave been unimaginable five years ago.Ah, progress!

I enjoy Seattle immensely, and see sev-eral former Brown CS folks on a regularbasis, including Dan Robbins ’91 (who’sin the office next to mine and one of myclose coworkers), Russell Belfer ’91, andMatt Ayers ’95/MSc ’98. And of coursethere are lots of Brown alums at Micro-soft: enough to fill a whole article withthe variety of jobs we have. Maybe in afuture conduit!?

John Gannon, one of Andy van Dam’sfirst and favorite students and one ofCS’s most distinguished graduates, diedin his sleep of cardiac arrest at his homein Silver Spring, MD, on June 12. Hewas 51.

John had a con-genital heart de-fect, but surgeryas a child hadmade it possiblefor him to live anactive and pro-ductive life. Hewas Chair of theDepartment ofComputer Sci-ence at the Uni-versity of Mary-land, where heearned a reputa-tion as a leadingresearcher insoftware engi-neering and,most impor-tantly to him, asa demanding butcaring teacherwhose door wasalways open.

Andy remembers John as exemplifyingthe best of Brown’s “get-involved-up-to-your-eyeballs” undergraduates: he was aUTA, a research assistant, paymaster (hekept the books for the graphics group),helped write proposals and critiqued ev-eryone’s work. John had a very idiosyn-cratic laugh (a wicked cackle) easilyevoked; he never took himself or otherstoo seriously, but he had a great serious-ness of purpose. He was always reliableand just got stuff done, no excuses—hehad a “My dog ate it” stamp made to useon the programs of students reduced tocreatively lame excuses for late or incom-plete submission.

At the time, Andy and Peter Wegner werethe two people teaching CS in the Divi-sion of Applied Math (John Savage wasteaching in the Division of Engineering)and Andy was trying to cover too manyaspects at once via AM101 and 102. Heincluded a little bit of parsing and com-piler theory, a subject then still in its in-fancy. One of the things John Gannon andAndy did was to try and make sense ofFrank DeRemer’s brand-new PhD thesison parsing to see what of it they couldteach to undergrads. John read the thesisand then would try to explain the basicsto Andy; Andy would find holes in the ex-planation, and back John would go to tryand work it out. After many such cyclesthey doped it out and managed to teachhot-off-the-press LR(k) parsing to meresophomores and juniors.

13

THE 23

Host Eugene Charn

Doing this digging had a life-long impacton John. He went to the University ofToronto for his PhD, as did many of hisgeneration from Brown, including EdLazowska, Frank Tompa, Chris Braun,Larry Weissman, David Elliott, JohnGuttag, and John Zahorjan—a pipelinefrom Brown to Toronto that continued formany years. In his professional career,John became a real authority on softwareengineering and brought rigor to an areasometimes lacking it. He also became asuperb teacher and mentor.

The John D. Gannon Scholarship Fundhas been established to commemorateJohn’s commitment to students and edu-

conduit

rd IPP SYMPOSIUM

cation at the University of Maryland. In-formation about this is available viahttp://www.cs.umd.edu gannon_memor-ial.html. Condolences may be sent toJohn’s family at this email address:gannon-memorial@cs.umd.edu. John, whowas raised in Rhode Island, is survived byhis brother, Rick, of Foster, and by hiswife, Nancy Garrison ’70, a Yale Lawgraduate who works at the U.S. Depart-ment of Justice. Nancy can be reached at10108 Day Avenue, Silver Spring, MD20910.

John’s many friends and colleagues atBrown mourn deeply his untimelypassing.

The Computer Science Departmentheld its 23rd Industrial PartnersProgram Symposium (less formallyknown as IPP day) on April 29, 1999,on ‘‘Web-based Natural-LanguageTechnology: Search, Translation andAnalysis.’’ As usual, John Savage,who runs the program, warmed upthe audience by having everyone in-troduce themselves and in the pro-cess demonstrated that he knew thenames of pretty much everyonethere (which, from my perspective, is

amazing feat). He then turned things

anover to me, and I promptly introduced thefirst speaker (whose name, fortunately, Idid know: Bill Woods).

Bill Woods is from Sun MicrosystemsLabs and has been doing work on usingnatural-language technology to improveinformation retrieval. His talk was enti-tled ‘‘Finding What You Really Want:Natural Language Technology in Preci-sion Content Retrieval.’’ The standardtechnology in this area is to start with theset of words in the user’s query and re-turn documents that contain those samewords. However, usually users do notwant an entire document, but rather justa passage that answers their question.This is the “Precision Content Retrieval”of the title, and it is much harder thandocument retrieval since the passage now

iak

sought contains many fewer words, andthus is less likely to contain exactly thewords in the query. Woods looked at howvarious natural-language technologiescan help in this problem. The techniqueshe considered ranged from the relativelybasic to those closer to the research edge.At the basic level I was surprised to learnthat ‘‘stemming’’ (reducing a word to itsbasic form—e.g., the word ‘‘reducing’’ be-comes ‘‘reduce’’) has a major positive im-pact in retrieval rates. Among the morecomplicated techniques described was‘‘subsumption’’—finding concepts that areeither more general or more specific thana given concept. The idea here is that ifthe user asks about, say, ‘‘computer’’prices, we might also want to give infor-mation about ‘‘workstation’’ prices, andvice versa.

The second talk of the day was by RoyByrd of IBM Watson, on ‘‘Text Mining forKnowledge Management.’’ Roy pointedout that one of the most profitable areasof IBM today is its consulting business.This business generates a lot of informa-tion about what IBM customers are doingand the problems they are having doingit. This information is both a problem andan opportunity. Essentially Roy and hisgroup at IBM are working on an on-linedocument creation and analysis packageto be used by the IBM consulting divisionto find documents that relate to their que-ries, route documents automatically to in-dividuals who could use the information,and use the entire document collection to

! 14

SymposiuTed Se

gather global information about trends inthe area. One of the technologies Roystressed was the use of finite-state au-tomata. Natural language is funny inthat few of the tasks involved can be com-pletely solved within the domain of finite-state machines, but a huge number ofthem can be approximated using thesetechniques. The advantage of doing so is

m speakers, clockwise from Eugene:lker, IBM; Steve DeRose, Brown; Bill

that finite-state technology is very robustand very efficient. Roy gave some num-bers on the speed of processing that werequite amazing. The talk also stressedthat the goal here was a system thatwould be available in the near future butwould, at the same time, incorporate avery large number of language tools. Theway to meet both of these desiderata is tokeep each tool fairly simple (which dove-tails with the use of finite-state tech-nology).

The first talk after lunch was given byTed Selker, also of IBM, and an IBMFellow. Ted is known, among otherthings, for inventing the track-point de-vice on IBM Thinkpads. Ted is also aBrown alumnus and took my AI course asan undergraduate. I invited him becausehe is well known as an interestingspeaker and he has done work that re-

Woods, Sun; Roy Byrd, IBM

conduit

lates to language issues. However, histalk today was not so much on languageas on questions about computer design inthe macro sense—if one is to walk aroundwith a computer, how big should it be,how should one hold it or otherwise carryit, etc. If the audience ever noticed thatthe talk did not relate all that closely tothe topic of the symposium I am sure theyquickly forgot—I know I did. This was theonly talk I have ever seen where thespeaker came equipped with fifteen totwenty computers and proceeded to pullthem out of his backpack and discuss thedesign issues involved in their creationand why they did or did not work as prod-ucts. One interesting point Ted made con-cerned computers used by people gather-ing interviewing data. It turns out that ifthe computer is designed so that only thepoller can see the screen, people are muchless willing to talk than if they can see thescreen at least part of the time. More gen-erally, Ted emphasized that one has tothink long and hard about how a com-puter is to be used in the real world beforecoming out with a product, and that mostof the machines he pulled from this back-pack never got produced commercially be-cause they failed this test.

The day’s penultimate talk was by StevenDeRose, who founded Electronic BooksTechnology (now Inso eBusiness Technol-ogies), but recently returned to Brownhalftime in the Scholarly TechnologyGroup. His talk was entitled ‘‘Links, Que-ries and Language Awareness on theWeb.’’ One way to think about this talk isthat whereas the other talks were aboutmaking computers better consumers oftext, this one was about making the com-puter’s job easier by having the humantext producer take on the burden ofmarking up the text to indicate some of itsmeaning. That is, we want marks that in-dicate not just how to display some text,but also indicate something about its con-tent. Steve first noted that hypertextmarkup language (HTML) goes a littleway in this direction, but not nearly as faras most people would like. The bulk of thetalk then was on XML (eXtensibleMarkup Language), which has markupsymbols for more content types and astandard way to extend the mark set toyour domain. One important point here—

! 15

I think Steve said it, but if not it was cer-tainly implicit in his talk—is that there isgoing to be a strong incentive to add thesemarkups to very many web documentproducers: the web search engines atsome point are going to be sensitive toXML markups, and thus if you want themto put your web site at the top of the listfor a user’s query, you had better put inmarkups that allow the search engine to‘‘realize’’ your site contains the informa-tion the user has requested (e.g., the priceof a new Ferrari).

The last talk of the day was mine, on ‘‘TheStatistical Revolution in Natural-Lan-guage Processing.’’ My goal here was notto illustrate how this technology can beused (I figured that the other speakerswould be able to do this far better than I),but simply to show how the use of statis-tics has revolutionized computational lin-guistics, and how it has led to remarkablyrobust and accurate programs that attacka wide variety of natural-language topics.In particular, I talked about four topics:parsing, word-sense disambiguation, pro-noun reference, and lexical semantics. Ifind the work in parsing (assigning a sen-tence structure to a string of words) par-ticularly exciting. As recently as six orseven years ago, there was no parser thatcould take, say, the front page of today’s

condui

A recent WiCS (Women in Computer Scieinvited to speak. WiCS acts as a resource fo

mentoring program and a ‘safe space

New York Times and parse every sentencetherein, even with numerous mistakes.Today we have several such programs (al-most all statistically based), and now theresearch effort is to drive down thenumber of errors the programs make. Theother topics in the talk were more se-mantic in nature. Even though statisticsare easiest to gather on surface phe-nomena like words, statistical techniquesare beginning to be applied to the lesssurfacy area of ‘‘meaning.’’ So in the workon pronoun reference we took some text,marked it up to indicate the referents ofall the pronouns, and then gathered sta-tistics on such things as the probability ofthe pronoun’s antecedent being, say, Nsentences back, for N = 0, 1, 2, etc. One in-teresting subproject here relates to thefact that pronouns in English have agender that must match the gender of theantecedent. Thus a goal was to automati-cally collect information about the typicalgender of objects described by words like‘‘piano’’ (neuter), ‘‘president’’ (mostlymale), etc. This project relates to the lex-ical-semantics portion of the talk, as thegoal of statistical lexical-semantics is tolearn semantic information about wordsby observing how they occur in text.

t! 16

nce) meeting at which Philip Klein wasr women in the department and offers a’ for women to discuss related issues

The astaff team is a blur of activity during a stuffing blitz to mail out ourgraduate recruitment posters. l to r: Fran Palazzo, Genie deGouveia,

Lori Agresti and the back of Dawn Nicholaus

NEW CS FACULTY

Vera, Emmanuel and ThomasHofmann at the CS picnic

The Department is verypleased to be welcomingthree new faculty mem-bers this year—WilliamChan, PhD University ofWashington, 1999; AmyGreenwald, PhD Cou-rant Institute, New YorkUniversity, 1999; andThomas Hofmann, PhDUniversity of Bonn,1997. Professors Chanand Greenwald won’t ar-rive until the new yearand will start teaching inSemester II, so newsabout them will be forth-coming in the springissue of conduit! Pro-fessor Hofmann is hereand teaching an ad-

ourse this semester on in-

vanced topics cformation retrieval and data mining.

Thomas and his wife Vera grew up in atown near Cologne. They met in highschool and discovered a mutual passion

conduit

for philosophy. During their time at theUniversity of Bonn, they also spent ayear in Paris, one of the great centers forcontemporary philosophy, which helpedcoalesce their future intellectual direc-tions. Despite having taken myriad phi-losophy courses, Thomas decided upon acareer in CS. Vera received her PhD inphilosophy. Their respective degrees wereawarded within days of each other andthey came to Boston, Thomas as apostdoc at MIT and Vera continuing herresearch at BU. Their son Emmanuel wasfive at the time and had to learn Englishfast. He now speaks English very well—he with an American accent, his parentswith an Oxford one. The challenge now isto make sure he retains his German.

From Boston, they headed for the westcoast and a year at Berkeley—Vera in theRhetoric department, Thomas in CS.They enjoyed Berkeley’s colorful ethnicdiversity, culture and climate. Fromthence they returned to the east coastand Brown. Vera now has a lectureshipin philosophy at URI and Emmanuel isnow in second grade in Barrington. Heespecially enjoys the school bus—he’s thefirst student picked up, so he’s aboard thelongest.

! 17

Thomas’ research goals can best be sum-marized by the motto “artificial intelli-gence by machine learning.” He isinterested in how computers can take ad-vantage of large amounts of data in orderto achieve better performance for a va-riety of AI-related tasks. His previouswork focused on problems in pattern rec-ognition, machine vision, information re-trieval, data mining, and natural lang-uage processing. Here, the term “data”spans a large spectrum from raw mea-surements and sensory inputs at one endto discrete and symbolic data at the otherextreme. Some of the key questions thatfuel Thomas’ re-

“One of his main motivations forcoming to Brown is its strong cultureof interdisciplinary research and its

opportunities for collaborationacross departmental boundaries”

search are: What arethe principles of in-ductive inference,i.e., how can we infergeneral laws from aset of examples?What are the mecha-nisms that would en-able machines tounderstand and interp

ret images? Howcan computers process and understandnatural language? What methods can bedevised to automatically detect structureand regularity in large data sets? Howcan data analysis and data visualizationbe combined to make human-computerinteraction more efficient? How can com-puters support humans in solving deci-sion problems in complex and uncertainenvironments? While he is greatly inter-ested in the mathematics and theoreticalfoundations of machine learning, he hasalso a profound interest in modeling andsolving specific problems. It is his strongbelief that the development of newmodels and methods can greatly profitfrom a confrontation with real-worldproblems, while on the other hand “thereis nothing more practical than a goodtheory” (V. Vapnik). The nature of theseproblems often requires a cross-disci-plinary approach, involving various as-pects of computer science from theory tocomputer graphics to systems design aswell as disciplines like statistics, infor-

conduit

mation theory, applied mathematics, sta-tistical physics, electrical engineering,operations research, and computationallinguistics. One of his main motivationsfor coming to Brown is its strong cultureof interdisciplinary research and its op-portunities for collaboration across de-partmental boundaries.

Philosophy is still a passion that oftenfinds Thomas reading and arguing intothe night. Vera is, of course, a worthy op-ponent in debate and there’s alwaysplenty to discuss. Thomas enjoys writingessays combining philosophy and com-

!

puter science, analyzing CS’s impact onsociety and the ensuing societal changes.Chess is another major interest. He playswith friends in Germany via the Internet,and thanks to an enrichment programrun by the California school system, Em-manuel is also able to play. Vera is an ac-complished pianist and church organist.She misses being able to play the organin her church back home to which shehad a key and could practice pretty muchat will. Fortunately for all, their piano re-cently arrived at their new home in onepiece.

As a former Green Party activist, Thomasavoids driving his car and opts insteadfor traveling by bike or public transportwhenever possible. He uses the East BayBike Path to come to work each morning(as does David Laidlaw). Since he’s nowin Rhode Island for the long haul, he willlikely become involved in politics again.

18

fac.activities@cs.brown.edu

The Manohar triplfrom Tom

Tom Doeppner. Tom enjoyed visitingSwami Manohar, PhD ’89, and his familylast January in Bangalore, India. He wasthere to teach a course on distributedcomputing for CS Industrial PartnerCompaq.

David Laidlaw. David’s new course oninterdisciplinary scientific visualizationis centered around writing mock grantproposals, reviewing them by emulatingthe NSF review process and recom-mending proposals for ‘funding.’ Prac-ticing what he preaches, David is co-PI ona new KDI grant from NSF, “3D Free-Form Models for Geometric Recovery andApplications to Archaeology,” with DavidCooper from Engineering as PI. He andSteve Reiss were just awarded an NSFgrant for visualizing program structureand execution to help in understandinghow large programs work (or don’twork...).

Franco Preparata. Franco was re-cently appointed chair of the Gödel PrizeCommittee. It is the most prestigiousaward in theoretical CS for an out-

standing paper(s)

▼▼▼

▼▼▼

ets sporting t-shirtsDoeppner

published in the pre-vious six years. InJuly he served on aninternational re-view committee forthe Department ofCS of the Univer-sity of Pisa, Italy. InAugust Franco pre-sented several lec-tures at a summerforum on supercom-puting organized atEl Escorial, Spain,by the UniversidadComplutense, inMadrid. Recently he

conduit!

was keynote speaker at a workshop intheoretical CS at the IBM T.J. WatsonResearch Center.

Steven Reiss. Along with David Laid-law, Steven has been awarded an NSFgrant for studying software visualiza-tion, as described elsewhere in this issue.He is also teaching CS233, where thissemester the class is attempting to builda modern Java programming environ-ment for the Suns. To keep him busybeyond these activities, he continues towork on providing tools for using designpatterns throughout the developmentprocess, designing user interfaces formobile computing, developing a front endfor searching the Internet with a localstartup company, Simpli.com, and har-vesting his vegetable garden.

Roberto Tamassia. Roberto gave akeynote lecture at the VIII Encuentros deGeometría Computacional in Castellón,Spain. His book, Data Structures and Al-gorithms in Java (coauthored withMichael Goodrich), has reached its 6thprinting.

Eli Upfal. In May Eli and PhD candi-date Gopal Pandurangan went to the31st ACM Symposium on Theory of Com-puting in Atlanta, where they presenteda joint paper. In July, Milos Hauskrecht(Eli’s postdoc) presented a joint paperwith Gopal and Eli at the 16th Interna-tional Joint Conference on Artificial In-telligence (IJCAI) in Stockholm. He wasappointed editor-in-chief of the Journalof Discrete Algorithms.

▼▼▼

▼▼▼

▼▼▼

▼▼▼

19

John Hughes

SIGGRABR

Andy van Dam. Andy co-chaired thefirst joint European Commission/Na-tional Science Foundation (EC/NSF) ad-vanced research workshop, entitled“Human-Centered Computing, OnlineCommunities and Virtual Environments,”held in France. As a member of the tech-

conduit

PH 99 SHOWCASESOWN WORK

This pof a

coveinuesates hSIGG

nical advisory board for Microsoft Re-search, Andy went to Beijing in June,where Microsoft has recently opened a re-search lab. Andy was part of a contingentincluding Raj Reddy (CMU), EdLazowska (Chairman of CS at the Uni-versity of Washington), and others thatprovided a full day of lectures on the fu-ture of computing to some 1500 invitedacademics and students—the first suchgathering in China. Andy’s talk was onPost-WIMP 3D User-Computer Interac-tion.

Stan Zdonik. In September Stan com-pleted his duties as North and SouthAmerican Program Chair for the 25th In-ternational Conference on Very LargeDatabases (VLDB) held in Edinburgh,Scotland.

▼▼▼

In LA this August about 50,000 peoplegathered at SIGGRAPH 99 to see what’snew in computer graphics. There were tu-torials, a spectacular “show floor” with allthe latest and greatest new products,panels, a “digital bayou,” a film and videoshow, panels on controversial topics,

icture, a computer-generated rendering Dr. Seuss-like scene, was chosen for ther of the SIGGRAPH Proceedings. This con-a trend: Brown students, faculty or gradu-ave been authors of papers featured onRAPH covers in four of the last five years

and technical paper sessions. At least25 people from Brown attended andparticipated in several venues. I’m notgoing to give the details of all ofthem—but I will describe the technicalpapers we presented, which is the areaI’m most familiar with.

PhD student Lee Markosian was leadauthor on one paper and coauthor onanother, both originating from hiswork on art-based modeling and ren-dering, to become his dissertation thisspring. The overall goal of this work isto incorporate into computer-generatedimages the kinds of skills that artistshave—representing complexity withjust a few strokes of a brush or pen,creating 3D shapes through “construc-tive drawing,” and selectively using de-tail to direct attention to what’simportant in an image. Lee’s SIG-

t

GRAPH papers were “(Gimme some)skin: A constructive approach to free-formmodeling” and “Art-based rendering offur, grass, and trees.” The first of these(by Lee, Jon Cohen ’00, J.D. Northrup ’00,and me) described how to build large andcomplex free-form shapes (like the torsoshown in Figure 2) by a technique basedon constructive drawing in which one de-fines a form by first placing large shapes

! 20

Figure 2: Jon Cmodeling syst

was assigned tone looked terand I told him t

got an anatoAnd again. L

short sleep bre

Figure 4: With tcreate the c

Durer-like imagused to conv

like cylinders or ellipsoids, cubes, andother “primitive” objects and then con-structing a “skin” around them. The skin,in this case, is an adaptively tesselatedmesh, positioned to have a user-chosenoffset from the underlying primitives inplaces where a primitive is nearby, and to

conduit

ohen, who had never used a commercialem and had no background in anatomy,o produce a torso with our system. His firstrible (it had no collarbones, for example),o try again. And then again. And then hemy text from the library, and tried again.ess than 24 hours after he started (with aak), he produced the model shown here

Figure 3: Atoon-style with crea

between theThis sort of mcannot eas

made with imsurfaces orconventio

modeling sohe same technology used toover image, we also makees, with dark and light strokes

ey highlights and rich texture

behave nicely elsewhere. “Nicely” meansthat it avoids self-intersection and triesto avoid wrinkles or creases, and degen-erate triangulations with long, thin, tri-angles. The speed with which one cancreate complex free-form models, espe-cially when one has Bob Zeleznik’sSKETCH system to help with placing theprimitives, is remarkable.

The other paper got to be informallyknown in the graphics group as “fuzzy furfeet,” because the first successful proof-of-concept picture was based on a drawingfrom Dr. Seuss’s The Foot Book. Thepaper was written by Michael Kowalski(BA ’98, Scm ’99), Lee Markosian, J.D.Northrup ’00, Lubo Bourdev (ScB, Scm’98), Ronen Barzel (ScB ’83), LoringHolden, and me. In an early draft of hispresentation, he spent the first few min-utes explaining why “non-photorealisticrendering” was useful. Fortunately, some-one who listened to it said “Michael: ifthey’re there to listen to you talk in thelast session of the last day of the confer-ence, I think you can assume that theybelieve what you did had a purpose!” I’mpleased to say that a lot of people werethere—about 1500 of them—and thatMichael’s presentation was a real suc-cess. Our work on this topic isn’t doneyet: Michael, now working at ATR inJapan, is collaborating with us on exten-sions to this work and another project onart-based graphics. Look for us at SIG-GRAPH next summer!

! 21

car-footsestoes.odel

ily beplicit

withnal

ftware

CHARNIAK UNPLUGGED

Eugene Charniak

Over the summer, President Gordon Geevisited General and Mrs. Kanellakis at their

home in Athens to thank them for establish-ing graduate fellowships in memory of theirson Paris and his family and for their support

of CS faculty. There was an exchange ofgifts—a gloriously illustrated book about

Greece from the Kanellakises and for them,a tie and a scarf specially designed as

Brown presentation gifts

Recently a newspaper article ap-peared in my e-mail on the difficultyuniversities are having in recruitingand retaining computer science fac-ulty. This got me to thinking about anumber of such articles I have seenover the last year.

The first of these was about a yearago, when high-tech companies werelobbying Congress to increase thenumber of special visas given to for-eign workers in areas in which therewas a shortage in the US. These com-

nies said that they were not able to hire

pahigh-tech workers, particularly in com-puter science and related areas, and thusit was important for the US to allow moresuch people to come into the US fromother countries. My thoughts on this weresomewhat ambivalent. On the one hand, Iam proud that the US, as the world’s pre-

conduit!

eminent multiethnic society, can drawthe best and brightest from around theworld. No doubt this is in part because ofthe salaries we pay, but I think fewwould disagree that an equally impor-tant reason is the extent to which ourculture is open to people of all races andethnicities. (Which is not to say we areperfect, just much better than most ev-eryone else.) Yet at the same time, I didnot like the idea of increasing the visaquotas. As a professor of computer sci-ence, I am disappointed that so fewAmerican students go into the sciences,and while I do not have any data on thissubject, my impression is that despitethe rising salaries high-tech workers cancommand, it is still the case that doctors,lawyers, and business executives out-earn computer scientists. If this is true,then the recalcitrant students are mak-ing reasonable decisions. Why studysomething hard, like science, when youcan major in something that is much lesswork and still earn more over your life-

22

With his f

...and one last g

I’d like toas an unMaster’s dtime whenas enrollmequipmencially gratdepartmethe warm

time?1 Since we live under the sway ofthe laws of supply and demand, not al-lowing more foreign workers in shouldcause salaries to rise even further and, Ihope, attract more US students into thesciences.

I was still mulling this over a few monthslater when I saw a second article, thistime about problems the military washaving in attracting enough new recruits.As this was a “news” article, the author

1. Our editor, Trina Avery, takes exceptionto the implication here that science isharder than non-science. Your humble au-thor stands by his prejudices.

conduit!

ather, Levie van Dam

limpse of Andyfest

Most o

thank Shuang Ji for his very genrestricted gift to the departmenegree working with Steve Reissthe department is experiencing cents soar, the number of faculty

t and space outstrips the adminisifying when a former student stent, and we are particularly gratefu sentiments with which it was give

Tom De

expressed no opinion on what should bedone about the problem, but a ratherlarge number of people were quoted whothought that the best solution would beto bring back the draft. They pointed outseveral benefits beyond the obvious one,that the military would not have toworry about where their next recruit wascoming from. One was that the youngwould have the experience of helpingtheir country in an important way. An-other that I remember particularly wasthat this would allow the civilian popula-tion to better understand military cul-ture, and would at the same time helpbring military culture into line with themores of the average civilian.

23

f the participants pose in front of theSciences Library

erous donation of $15,000t. Shuang completed his

in ’94. His gift comes at aonsiderable growing pains

increases, and the need fortration’s largesse. It’s espe-ps forward to help out thel to Shuang for his gift andn.

an, Chairman

Department of ComputeBrown Universit

Box 1910, Providence, RI

conduit!

Printed on recyled pape

Th

My first reaction was that the author andthe quotees were on to something impor-tant here, but that they were not think-ing big enough. Here, I thought, was thesolution to the computer programmershortage as well. We should draft peopleinto Microsoft, Sun, Compact, SGI, Intel,etc. Not only would this solve the pro-grammer shortage, it would give theyoung a good feeling for having helpedthe country surmount a national prob-lem, it would help the civilian world un-derstand the nerd culture, and finally, itmight even bring the culture inside of(say) Microsoft or Sun in line with that ofthe rest of the country.

It was with all this baggage rumblingaround in my head (please excuse themixed metaphor, but it best describes mystate of mind) that I read the article on

r Sciencey02912, USA

conduit

r

conduit!A publication of

e Computer Science DepartmentBrown University

❦Inquiries to: conduit!

Department of Computer ScienceBox 1910, Brown University

Providence, RI 02912FAX: 401-863-7657

PHONE: 401-863-7610EMAIL: sjh@cs.brown.edu

WWW: http://www.cs.brown.edu/publications/conduit/

Suzi HoweEditor-in-Chief

the shortage of computer science profes-sors. As I saw it from my previous read-ing, there were three ways to attack thisproblem. Increasing the quotas for for-eign professors of computer sciencewould probably not do any good. Frommy experience here at Brown, there doesnot seem to be any effective quota here.The typical person we wish to hire hasunique credentials, a case can be easilymade that there is no person in the USwith the same credentials, and, ... well,you get the general drift. This left theother two options: raise the salaries ofprofessors of computer science, or use thedraft. On this issue, as you might im-agine, I have very firm opinions.

! 24

NON-PROFITU.S. Postage

PAIDProvidence, RI

Permit #202

Address changes welcomed

Jeff CoadyTechnical Support

Katrina AveryEditor