1995 Gordon Bell Prize Alan H. Karp Hewlett-PackardLaboratories Michael H e a t h Universityof Illinois
Al Geist Oak Ridge National Laboratory
m Although performance gains were not as impressive as last year's, one trend continued: The highest performance came from the big iron. 0018-9162/96/$5 00 0 1996 IEEE
T
he Gordon Bell Prize recognizes significant achievements in the application of supercomputers to scientific and engineering problems. In 1995, finalists were named for work in two categories: performance, which recognizes those who solved a real problem in less elapsed time than anyone else, and price/performance, which encourages the development of cost-effectivesupercomputing. One entry was received in the third category of compiler-generated speedup, which measures how well compiler writers are doing at easing the programmiingof parallel processors. In addition, we honored another group in a category that no longer exists, special-purpose machines. Gordon Bell,who recentlyjoined Microsoft to develop video and highspeed networking technology, is sponsoring $2,000 in prizes each year for 10 years to promote practical parallel processing research. In light of the modest gains in price and price/performance over last year's spectacular gains, the total amount of prize money has been reduced for this year. This is the eighth year of the prize, which Computer administers. The winners were announced December 7 at the Supercomputing 95 conference in San Diego.
RESULTS Panayotis Sltordos of the Massachusetts Institute of Technology was recognized wil h a $500 award for his modeling of air flow in flue pipes. He achieved 3.6 Gf/s/$M (billion floating-point operations per second per million dollars) on a cluster of 20 Hewlett-Packard workstations. Masahiro Yoshida, Masahiro Fukuda, and Takashi Nakamura of the National Aerospace Laboratory in Japan, Atushi Nakamura of Yamagata University, and Shinji Hioki of Hiroshima University submitted a quantum chromodynamics simulation run on 128processors of the Numerical Wind Tunnel. 'Their original submission reported a computation rate of 112 Gf/s, lower than last year's winner. However, by August of this year they had improved to 179 Gf/s, for which they received $500. One entry was received in a category that no longer exists in the Bell Prize competition, special-purpose machines. However, the results presented by Junichiro Makino and Makoto Talji impressed us enough to recommend a $250 award. A simulation of the motion of over 100,000stars was run at 112 Gf/s on a machine with 288 computation engines having a peak performance of 176 Gf/s. A wide variety of applications were submitted by the remaining six entrants, one ofwhom submitted two different applications.Agroup of 16 January 1996
people from the Center for Supercomputing Research and Developmentat the University of Illinoissubmitted the parallelization of two large applications by Polaris, a parallelizer that converts sequential Fortran to SGI Parallel Fortran. Albert Greenberg and R. Srikant showed how to predict network performance using message passing on an SGI multiprocessor. Entering from Australiawere Marek Michalewicz of the Commonwealth Scientific and Industrial Research Organization and Roger Brown of Cray Research. They computed the electronic structure of a system containing more than 500,000 atoms, one of the largest quantum mechanical simulations yet undertaken. We received two entries from Stephen Taylor of the California Institute of Technology, both involvingsimulations of space vehicles. Another aerospace application was submitted by J. Vadyak, G.D. Shrewsbury, and K. van de Houten of the Lockheed-Martin Skunk Works, Gary Montry of Southwest Software, and six people from Intel Scalable Systems Division. One enterprising entrant submitted an application running at a very impressive price/performance of almost 8 Gf/s/$M. Unfortunately, the runs were made on a single processor, which doesn’t qualify for the Bell Prize, even if the processor supports instruction-level parallelism. Although the performance gains were not as impressive as last year’s, one trend did continue: The highest performance came from the big iron. In fact, only one submission used a cluster of workstations. This entry won an award for price/performance, but it lagged by a factor of over 1,000 the best raw performance reported this year. In light of the modest gains, which contrast with last year’s spectacular improvements,the size of the cash awards was reduced.
flow, which follows the motion of the fluid. In spite of the difficultyof followingthe evolution of turbulence and even shock fronts, the time evolution of the flow is relatively slow. Such a pace allows the use of relatively large time steps in the modeling. The sound of a whistle, on the other hand, depends on acoustic waves, which move at the speed of sound. Following such fast-moving structures requires very short time steps. Until recently, the amount of computation per time step and the number of time steps required made such models impractical. Actually, there is one advantage in being required to take such very small time steps. Hydrodynamic models typically use complicated implicit methods to allow the use of long time steps. This year’s awardee for price/performance, Panayotis Skordos of MIT, was able to use a simpler and more easily parallelized explicit method. Explicit methods use information from only a few (sometimes only one) recent time steps to determine how much the density andvelocity change in going to the next time step. Since the information used is local, any movement that crosses more than one grid point between updates will lead to inaccuracies. This limit is formalized in the Courant condition, which states that the solution computed by an explicit method will be unstable unless the time step is less than the time it takes sound to travel from one grid point to another. Hence, fine grids require short time steps. Implicit methods avoid the instability inherent in the explicit methods by including more information. In particular, these methods express the new values of density andvelocity as a function of their current values and their values at the new step. This circular reasoning means that implicit methods need to solve a system of equations to determine the new values. In exchangefor this added complexity, the time step is limited only by the rate of change of the flow, not the relation between the sound speed and grid size. In practice, implicit methods allow time steps 10 to 1,000 times larger than the Courant condition limit. Implicit methods are hard to parallelize because of the Think about blowing a whistle. Blow softly and you get global nature of the solution of the system of equations. one sound; blow harder, you get a different sound. Blow Basically, every grid point affects thevalues of every other very hard and the pitch changes dramatically. Aside from grid point at the next time step. the World Famous Whistle Band, you might ask, who Explicit methods parallelize much more easily. Say we cares? Musicians, for one. A whistle is a simple example partition the grid into blocks, one block per processor. of a variety of wind instruments, everything from an organ Each processor can update the points in the interior of its pipe to a recorder to a flute. Understanding the require- block without communication. Only when it needs to ments for producing a pleasing sound from these instru- update its boundary does the processor need more data, ments is of great interest. and even then the data comes from a neighbor. As the Actually, we know very little about why these instru- blocks get larger, the ratio of computation (which is proments sound the way they do; the work reported here is portional to the area of the block) to communication the first comparison between direct simulation and exper- (which is proportional to the perimeter) increases. Hence, iment. In contrast, string instruments such as violins and the fraction of the time spent communicating decreases. guitars have received a lot of attention. The reason goes The problems run for this entry had so much computabeyond the million-dollar prices paid for old instruments. tion per communication that even with a slow Ethernet connection and all the overhead associated with TCP/IP, It’s much harder to model a whistle than a Stradivarius. It would seem that we know how to model the flow of communications were tolerable. Typically, less than 20 air. After all, the aircraft industry has been modeling the percent of the elapsed time was spent waiting for data. behavior of aircraft since even before the advent of com- More recently, Skordos reduced this overhead to 10 perputers. However, modeling the lift of an aircraft’s wings cent by the simple trick of computing the boundary points is a piece of cake compared with figuring out the sound and sending them first. With this simple change, commuproduced by a whistle. nication occurs while the interior points are being The airplane’sbehavior is determined by hydrodynamic updated. Computer
Skordos studied the behavior of a very simple form of a whistle. His whistle, actually a model of a simple recorder, is a 20-cm-longcylinder. At one end is the mouthpiece and a flue channel, which guides the blown air to a sharp edge, called the labium, and into a closed pipe, which acts as a resonant cavity. The labium causes acousticwaves to form in the incoming air; these propagate down the pipe, bounce back, and interact in a complex way with the inflow. The timescaleis quite short; it takes only about a millisecond for the acoustic waves to reflect back to the labium. To make the problem tractable, Skordosmodeled the system in two dimensions. As anyone who has played the recorderwill know, this approximation is not exact, because the labium is flat while the tube is cylindrical. Even with this approximation, the models still require a lot of grid points, anywhere from 400,000 to 1.5 million. With the power of 25 clusteredHP workstations(20 in use at anyone time), running at almost 180 Mf/s, it took more than two days to compute the 150,000 time steps needed to model about 30 milliseconds of air flow. Fortunately,this simulation is just long enough to predict the spectrum. The results compare quite well with direct measurements, especially when we consider the simplifymgapproximations. Skordos’s program ran on a collection of 25 machines that were being used by other people. To accommodate the realities of working in such an environment, he built his runtime system to allow processes to move from one processorto another. Each such migration took 30-90 seconds, but he arranged things so that they occurred only once or twice per hour. Checkpointingis another problem, since all the processors writing their data at once would use up the entire file server for a considerable length of time. Rather than letting all the tasks checkpoint simultaneously, he required them to proceed one at a time. While this choice meant that each task took three times as long to checkpoint its data, other people were able to use the file server during this time. Since checkpointing tookless than 2 percent of the total runtime, he could get good performance without having to hide from his colleagues.Even with the poor network and the need to consider other users, Skordos achieved 3.6 Gf/s/$M, almost 40 percent better than last year’s winner. One interesting aspect of Skordos’swork is that each processor can be working on a different time step. This approach is good, since any time you can avoid the need for all processors to agree on something, your parallel job will run faster. However, most simulationsrun in lockstep; at each time step they wait for all processors to complete their work before moving ahead. It’s hard to do anything else with an implicit method. Skordos is using an explicit method, so each processor needs to share data only with the processors holding grid points adjacent to the grid points on the boundary of its block. In this case, we can have processor 1send its Values for time N to processor 2. Processor 2 can compute its values at time N 1 and send them to processor 3. Processor 3 updates its points to timeN 2, and so forth. If for some reason processor 1has been delayed, it might still be working on updating interior points at time N. This flexibility means that there is no need for any global com-
+
+
munication. Even at one of the periodic checkpoints, the interaction among the processors is limited to synchronizing to ensure a consistent state.
PERFORMIANCE The search for the essential nature of matter, the set of elementary particles, has a long history. The ancient Greeks had four: earth, water, air, and fire. Later, we had a large number, the atomic elements. By the middle of the 20th century, we were back to three: the electron, the proton, and the neutron. By 1960, we were back to a large number-a whole zoo of pions, kaons, muons, and so forth. Next, it was back to a small number. How many readers remember the discoveryof the omega minus particle that filled in the “eightfoldway”? The most widely accepted theory today also has a modest number of elementary particles divided into two groups. The light particles-the leptons-are represented by the electroin and include the neutrino and the muon. The heavy pal-ticles-the baryons-exemplified by the proton, but including a very large number of other particles, are now thought to be made up of just three fundamental particles, the quarks. Quarks are unusual in a number of ways. Unlike other particles, they have charges that are a fraction of the charge on the electron. A proton consists of two quarks with charge 2/3 and one with charge -1/3; a neutron consists of two quarks of charge -1/3 and one with charge 2/3. Other particles are made up of other combinations of quarks. For example, a neutral kaon is made up of a quarkwith charge 1/3 and one with charge -1/3. Perhaps the most unusual feature of quarks is that they are never found in isolation. The current explanation is that the force that binds quarks together gets stronger the farther apart they get. In contrast, quarks that are sufficiently close can hardly feel each other’s presence. Contrast this behavior with other forces, such as gravity or the electromagneticforce on charged particles; these get weaker with increasing distance. Even the forcle that binds protons in the nucleus weakens beyond some critical distance. Particle physicists want to understand why the force between quarks is different. The theory used to explain quark behavior is called quantum chromodynamics (QCD) because it is similar in many ways to what is arguably the most successful theory in the history of physics, quantum electrodynamics (QED), which explains the behavior of electrons. The basic idea of QED is that electrons interact with each other by exchanging another particle, a photon. Hence, two electrons repel each other by exchanging photons that push the electrons apart. QCD is more complicated than QED. Quarks still interact by exchanging a particle, in this case a gluon. However, quarks and gluons are more complicated than electrons and photons. An electron is most likely a point particle having a single property, a spin direction; a photon is massless, so it travels at the speed of light. Quarks, on the other hand, have structure, and gluons act as if they have mass. January 1996
The structure of a quark comes about because the quark can be in one of several states. These states are denoted by the term color, hence quantum chromodynamics. The upshot of all this for theoreticians is that an electron can be represented by a single, real number, and its interactionwith a photon involves nothing more than multiplications. A quark is described by a three-element vector of complexnumbers. When it interacts with a gluon, its vector components change; a physicist would say its state has changed. This change from one state to another is represented bymultiplying the three-element vector of the quark by a 3 x 3 complex matrix representing the effect of the gluon. The goal of the work submitted by M. Yoshida, M. Fukuda, and T. Nakamura of the National Aerospace Laboratory, A. Nakamura of Yamagata University, and S. Hioki of Hiroshima University is to understand how quark confinement works by studying the long-range behavior of gluons. The models were run on the Numerical Wind Tunnel, a machine consisting of 140vector processors each having a peak performance of 1.7 Gf/s. These processors are connected by a crossbar switch capable of transmitting over 800 megabytes per second between processors. As its name indicates, the machine was designed for computational fluid dynamics problems. On a very different problem, this group managed to achieve 180 Gf/s, over 80 percent of the peakperformance of the 128 processors used. QCD is most often modeled with a lattice gauge model in which four-dimensional space-time is divided into a finite set of grid points connected by links. Each link has associated with it a set of 3 x 3 complex matrices called the linkvariables. For each grid point, denoted by its space coordinates and time, there is a unique matrix for each coordinate direction in the grid. These matrices contain all the information on the gluons Physical systems tend to follow the path of least action. (The principle is easily understood by watching a child doing chores.) Basically, a collectionof parhcles will minimize the amount of energy expended over time. The value of this integral, called the action, is presenred on the lattice under a certain set of coordinatechanges called gauge transformations. Hence, the name-lattice gauge models. The problem of minimizing a very complicated function, the action, over space-time is handled with a Monte Carlo approximation. Basically, a simple perturbation is made to the system, and the action is recomputed. If the action is reduced, the change is accepted. If the action is not reduced, the change is accepted with some probability that falls off rapidly as the difference from the old action increases. Clearly, it takes a lot of steps to reach a solution, and each step involves a lot of work. The model submitted uses a grid of 64 points in each of the three space dimensions and 63 points in time. (Note the rare, correct use of the term “point in time.”) Each of these 16 million points is represented by four complex, 3 x 3 matrices. Each energycalculation, the core part of the algorithm, involves four products of these matrices. More products are needed to compute the gluon properties. On Computer
average, more than 5,700 floating-point operations are needed for a single update to a grid point. Since each grid point must be updated many thousands of times during a run, performance is critical. There are some subtleties to the parallelization. The data partitioning is key. Because the processors have vector hardware, two of the space dimensions were kept together. The question was how to partition the remaining space dimension and time. A square has a smaller perimeter than a thin slab, so there will be less data communication with the former. On the other hand, a slab needs to get data from onlytwo neighbors; a square decomposition needs data from eight processors. The first implementation was a decomposition into nearly square rectangles. Each processor computed its part of the grid independently, synchronized with its eight neighbors (including neighbors on the corners) to get the latest values of the points next to its local boundary, then started the next step. After considerable tuning, the program ran at 140Gf/s. Alater implementation decomposed the two dimensionsinto slabs. Even though more data was transferred, the program ran at 180 Gf/s because synchronization was needed with only two neighbors. In fact, the slab implementation was able to overlap almost all the communication with computation. This work has already produced a significant result. QCD does not specify a mass for gluons. However, through their interactions they can act as if they have mass. The simulations reported here were able to show that this effective mass depends on the distance traveled. Further work may be able to show that quark confinement is a result of the additional energy required by the increasing effective mass of the gluons exchanged by quarks as they move apart.
SPECIAL-PURPOSE MACHINES Black holes are strange objects. They are so massive that nothing, not even light, can escape if it gets within a certain distance of the object, a surface called the event horizon. An object approaching the event horizon experiences the same kind of relativistic time dilation observed when bodies move past each other at high speed. Theory says that within the event horizon, time and space exchange roles; objects can move forward and backward in time, but only forward toward the center in space. Recent observations have skown that the centers of galaxies are also strange objects. Some emit most of their energy in X rays. Others are highly variable. High-resolution photos from the repaired Hubble Space Telescope show small, complex structures. Spectra of these regions show stars moving at extremely high velocities. What is a poor astrophysicistto do when faced with two bizarre kinds of objects?Explain one in terms of the other, of course. The entry submitted by Junichiro Makino and Makoto Taiji of the University of Tokyo was awarded a prize of $250 for the performance of a simulationof galaxy cores with black holes having a mass a million or more times that of the sun. There are three basic kinds of galaxies.The spiral galaxies like our own are the most familiar, having a central bulge and a big disk of stars, dust, and gas. The second group, the ellipticals, are ellipsoidal in shape and have very
little dust. The class of irregular galaxiesis basicallyacatchall for galaxiesthat are neither spirals nor ellipticals. Theoristshave tried for many years to explain the structure of elliptical galaxies. Models that attempt to form them out of collapsing clouds of intergalactic gas invariably end up with highly flattened structures that rotate much too fast. Elliptical galaxies can’t have evolved out of spiral galaxiesthat have used up all their dust because the ellipticals have so much more mass. It wasn’t until 1977that aviable proposal appeared. Alar Toomre suggested that ellipticals formed from mergers with other galaxies. This idea explained severalfeatures of ellipticals, such as their density distribution, the velocity dispersion (the range of velocities of the stars), and the ratio of the maximum rotation velocity to the dispersion. The models did not do such a good job of describing how the cores of elliptical galaxies formed. If collisions of galaxies without black holes can’t produce the kind of elliptical galaxieswe observe, and if many galaxieshave super-massiveblack holes at their cores, it is logical to ask what happens if two such galaxies collide. The result is a very complicated modeling problem. There are severalways to model a systemwith two massive objects and a large number of much smaller ones. A restricted three-body problem makes the assumption that the gravitational effectsof the small bodies are negligible. However, there are a lot of stars, so their cumulative effect can be large. Also, two stars often get so close to each other that their mutual attraction outweighs the effect of the black holes, at least for a short while. More detailed models need lots of computer power because the gravitational force extends to infinity. Hence, every star affects the motion of every other star. A straightforward calculationof the forces takes a time proportional to the square of the number of stars. Recent work, including that of a 1992 Bell Prize winner, has focused on methods whereby the time to compute the forces increasesmore slowly. Unfortunately,these more efficient methods don’t handle problems like the one studied in this entryverywell. The faster methods approximate a collection of stars distant from the one of interest as if the collection were a single, larger body. This approach works well when the stars have a relativelyuniform distribution. When there is a very strong concentration of stars near some point, as in the center of a galaxy, most of the calculation ends up being done the old-fashioned way, star by star. The appropriate step size used to advance the stars to their new positions is also a problem. The motion of each star is approximated by segments of a fifth-order polynomial. If the stars are far apart, these segments are a good approximation to the motion and can be rather long; if the stars get too close, the segments must be shorter because they don’t approximate the paths of the stars as well. In the center of a galaxy, the step sizes of different stars vary by up to 11orders of magnitude at any time. It is clearly impracticalto follow all the stars at the shortest time step, yet using a larger time step leads to inaccurate results. The onlyviable approach is to use a different time step for each star, one that doesn’t fit in with the faster algorithms. Faced with the problem of simulating the motion of a large number of stars with a slow method, this year’s honorees decided to build a special-purpose machine. Since
most of the tiine in the simulation goes into calculating the forces between pairs of particles, they built a parallel machine made up of chips optimized to do the force calculation. The Grape-4 (for gravity pipe) system will ultimately consist of a ho,stinterface, a control board, and 40 processor boards. Each processor board holds 48 custom LSI chips that callculate the gravitational force and its first derivative. Since each chip is capable of computing at over 0.6 Gf/s, the total system of 1,920 processors has a peak performance of almost 1.3 Tf/s (trillion floating-point operations per second). The calculation submitted used only 288 processors on six boards to simulate the motion of over 130,000 stars orbiting two black holes. The simulation ran at 112 Gf/s, almost 60 percent of the peak rate. The simulations can help in answering a number of questions. First of all, the density distribution in the models more nearly matches the distribution actuallyobserved than does the density distribution in models without black holes. A more interesting question is that of multiple mergers. Since elliptical galaxies are up to 100 times as massive as spirals, each must have swallowed several other galaxies. If each has a black hole at the center,what is the appropriate model‘) Do the black holes from one merger coalesce before the next merger or do they stay separate? If they stay separate, models might have to include an arbitrary number of black holes. Fortunately, the work done on the Grape-4 machine indicates that the black holes merge in a relativelyshort time, so that each merger can be modeled with only two massive objects.
OTHER ENITRANTS Seven other entries submitted by six teams presented some challenging work. Each of these entries demonstrates an advance in the state of the art. A group from the Center for Supercomputing Research and Development submitted compilationsof two complete applications with a new parallelizing compiler.This compiler takes Fortran 77 source and adds parallelization directives for the Silicon Graphics Challenge multiprocessor. Om two applications they demonstrated speedup far superior to that of the parallelizing compiler supplied with their SGI Power Challenge with eight processors.Thle first models turbulent fluid flow with strong shocks; the second is used to model atmospheric convection. They also showed good speedups on a wide variety of applications taken from the Perfect Club Benchmarks and SPEC95fp. Gerhard Fasol of the University of Tokyo made two interesting contributions. First, he developed ai numerical method for solving the time-dependent Schrodinger equation for a system of electrons moving in a strong magnetic field. His solution takes only a time proportional to the number of grid points; other methods either make simplifying approximations or take a time proportional to the cube of the number of grid points. He was able to solve problems with up to 3 million grid points. His second contribution was to carefully structure the memory accesses of an application to achieve Gp to 140 Mf/s on a 99-MHz January 1996
HP workstation having a peak performance of 200 Mf/s. Even on large problems, his code ran at almost 90 Mf/s. Since Albert Greenberg and R. Srikant are from Bell Labs, it isn’t surprising that they studied the performance of communications networks. They developed a new method for studying complex networks carrying traffic with different properties, for example, data and video. Ordinarily, these models involve direct simulation of all the links, assuming random inputs. The new method reduces this global problem to one of solving a large number (60,000 in the problem submitted) of two-node problems. While the time to solve the global problem grows exponentiallyas the network gets larger, the solution time with the new method increasesmuch more slowly. In addition, these smaller problems can be solved in parallel. The entry submitted by Marek Michalewicz and Roger Brown of the Commonwealth Scientific and Industrial
~
computer
Research Organization in Australia computes the electronic strumre of a solid. Unlike earlier studies, which assumed periodic boundary conditions suitable for crystals, their work applies to disordered materials. They solved the Schrodinger equation to find the energy levels of a system consisting of over 550,000 atoms. They can follow such large systemsbecause the computer time required increases only linearly with the number of particles. Equally important, their algorithm is parallelizable and vectorizable, so they runa Cray C90 at over half its peak speed. Stephen Taylor of the CaliforniaInstitute of Technology sent in two submissions that demonstrate the progress being made on real-world problems. One is a simulation of exhaust from a rocket. This particle-in-cell code followed over 35 million xenon ions, including the effects of the electric charges. Such simulations are important in determining how much the backflow from the exhaust
affects the rest of the spacecraft. His second entry computes the buffetingexperienced by a rocket during launch. These vibrations can affect payload instruments as well as the guidance system. A simulation with almost a million grid points required about 120 hours on an Intel Delta machine with 256 processors, and larger models are needed to get sufficientlyaccurate results. A team from the Lockheed-Martin Skunk Works also looked at aerospace problems. They were able to model a complete fighter aircraft with 22 million grid points on an Intel Paragon with 1,024 nodes. The model included the effects of the rotating propellersof this ASTOVL (advanced short takeoff and vertical landing) aircraft while hovering and flying. The second problem involves calculating the radar reflectivity of aircraft. High-frequency radar, over 1 GHz, can be modeled by approximate means, but low frequency, 100 MHz to 1GHz, requires the solution of Maxwell's equations. The discretization requires the solution of a complex matrix with up to 75,000 rows. The solution of a single problem of this size took over 240 hours on an iPSC/860 with 128 nodes. I
THE JUDGES A l a n H. K a r p , who chaired the judging committee, is a senior member of the technicalstaff at Hewlett-Packard Laboratories in Palo Alto, Calijornia. Michael H e a t h is a professor in the Department of Computer Science and a senior researchscientist at the National Centerfor SupercomputingApp1ication.sat the Universityof Illinois in Urbana-Champaign. A l G e k t is asenior researchscientistat OakRidge National Laboratory and leader of the computer science group there. Karp can be contacted at Hewlett-Packard, 1501 Page Mi// Rd., PaloAlto, CA 94304, e-mail
[email protected]. Heath's address is Department of Computer Science, 2304 Digital Computer Laboratory, University of Illinois, 1304 West Springfield Ave., Urbana, IL 61801 -2987 e-mail heath@ ncsa.uiuc.edu.Geist is at Oak Ridge NationalLaboratory, PO Box2008,Oak Ridge, TN 37831 -636i: e-mail geist@msr. epm.ornl.gov.
Call for Papers WET ICE '96 "Collaborating on the Internet: The World-Wide Web and Beyond" June 19-21,1996 :Stanford University, California, USA Co-sponsors IEEE Computer Society and the Concurrent Engineering Research Center (CERC) at West Virginia University are accepting papers for the Fifth Workshops on Enabling Technologies: Infrastructure for Collaborative Enterprises (WET ICE '96). The event will be hosted by Stanford University's Center for Design Research (CDR) and coordinated by CERC. We are applying for the cooperation of ACM SIGOIS. Papers reporting survey, original research, design and development, and applications of enabling technologies for collaboration, particularly technologies that utilize the Internet, are sought for the following proposed workshops: Web Infrastructure for Collaborative Applications Shared Design and Prototyping Environments Project Coordination Electronic Notebooks Distance Learning Papers accepted for the workshop will be included in the postproceedings to be published by IEEE Computer Society Press. PostScript copy of the full paper due Notification of decisions Workshop Final version of the paper due
March 15, 1996 April 19, 1996 June 19-21, 1996 June 28,1996
For more information about the individual workshops and about paper submission, see the WET ICE '96 homepage at
http:I'/www.cerc.wvu.edu/WETICE/ WETICE96.html General Chair: Charles Petrie, Stanford University Program Chair: Srinivas Kankanahalli, West Virginia University Local Arrangements Chair: Vinay Kumar, M/Cast Commun., Inc. Program Committee: Mark Klein, Penn State U. Hussein Abdel-Wahab, Old Richard Kouzes, West Virginia Dominion tJ. Takashi Aranos, NTT Corp. U. Robert Marcus, American Chandrajit Bajaj, Purdue U. Management Sysems., Inc. Gene Bouchard, Lockheed Martin John Callahan.,West Virginia U. Feniosky Pena-Mora, MIT Paolo Ciancarini, U. of Bologne Alexander Schill, Technical U. of Dresden Prasun Dewan, U. of North Richard Mark Soley, Object Carolina - Chapel Hill Taher Elgamal, Netscape Management Group P. David Stotts, U. of North Communications Corp. Carolina - Chapel Hill Sumit Ghosh, .Arizona State U. C. E. Thomas, Oak Ridge K. Gopinath, Indian Institute of Science National Laboratory Giuseppe Iazeolla, U. of Rome
I
I
Point of Contact: Srinivas Kankanahalli CERC, P.O. Box 6506, West Virginia U. Morgantown, WV 26506 F'hone: 304-293-7226, Fax: 304-293-7541 Email:
[email protected]
I
I
