Keynote Speaker: Henri E. Bal

HENRI E. BAL received a M.Sc. in mathematics from the Delft University of Technology in 1982 and a Ph.D. in Computer Science from the Vrije Universiteit in Amsterdam in 1989. His research interests include parallel and distributed programming, cluster computing, metacomputing, interactive applications, and programming languages. He supervised several projects in these areas, including the Orca parallel language, the Panda/LFC communication substrate, the Manta high-performance Java system, and the Albatross distributed supercomputing project. He also is one of the key researchers in the Distributed ASCI Supercomputer (DAS) project of the ASCI research school.

At present, dr. Bal is a full professor at the Faculty of Sciences of the Vrije Universiteit, where he heads researchs groups on parallel programming and Physics-applied computer science. He is author of "Programming Distributed Systems" (Prentice Hall, 1991) and coauthor of "Programming Language Essentials" (Addison-Wesley, 1994) and "Modern Compiler Design" (Wiley, 2000). He is a member of the IEEE CS European Distinguished Visitor's Programme.

Programming Support for Distributed Clustercomputing
Henri E. Bal
Department of Mathematics and Computer Science
Department of Physics and Astronomy
Faculty of Sciences, Vrije Universiteit, Amsterdam, The Netherlands

There is a resemblance between the evolution of sequential machines from mainframes to minicomputers to PCs and the more recent evolution of parallel machines from supercomputers to clustercomputers. Mainframes and supercomputers typically have many users and therefore are ill-suited for interactive applications. Minicomputers and clusters are cheaper and have far fewer users, allowing more interactive use. Personal Computers are used primarily for interactive tasks, but idle PCs can be combined to run computation-intensive parallel programs. The parallel equivalent of a PC, a Personal Clustercomputer, may become a reality with further progress in system integration.

As clusters become more widespread and more dedicated to small groups of users, two new types of applications can be foreseen. First, a single (local) cluster can be used to run interactive applications, which typically need a substantial amount of dedicated compute power to obtain real-time responses. Clusters used in this way, however, will also be idle a large fraction of the time (just as PCs are), so a huge amount of unused computing capacity will exist worldwide. Another way of using clusters therefore is to run grand-challenge applications on multiple idle clusters at different geographic locations. This idea is similar to distributed supercomputing, except that the distributed computing platform is hierarchical: it contains fast intra-cluster networks and slow inter-cluster networks.

The Advanced School for Computing and Imaging (ASCI, The ASCI research school is unrelated to, and came into existence before, the Accelerated Strategic Computing Initiative) has built a prototype of such a hierarchical distributed supercomputer, called DAS (Distributed ASCI Supercomputer, Papers about DAS, CAVEstudy, Albatross, and MagPIe are available from http://www.cs.vu.nl/~bal/) and is in the process of building a successor, DAS-2. DAS-2 will consist of five SMP-based clusters located at five Dutch universities and will have about 300 processors. Each university can use its own cluster for interactive applications; the entire distributed system can be used for very demanding computations.

Interactive applications allow the user control over an ongoing computation (usually a parallel simulation running on a cluster). By using virtual reality environments (such as a CAVE or a tiled video wall) to visualize the output of the simulation, this interaction can be done at a high level of abstraction. We have built a toolkit called CAVEstudy with which such applications can be built. CAVEstudy has been used to steer the simulation of molecules, chaotic systems, and robot-soccer. A user in a CAVE can trigger parallel computations by manipulating virtual objects in the natural domain of the application (e.g., moving an atom, kicking a ball).

Distributed supercomputing applications can be run on multiple clusters connected by a wide-area network (WAN). One would expect that the relatively high latency and low bandwidth of WANs severely limits the type of applications that can be run efficiently in this way (SETI@home for example uses independent tasks that each take one day). In the Albatross project, we study how applications can be optimized for multiple clusters, by taking the hierarchical structure of the system into account. By reducing or hiding the wide-area communication overhead, several more medium-grained applications can run efficiently on wide-area clusters. In our current research, we are developing programming environments that do some of these optimizations automatically. MagPIe, for example, is an MPI library that optimizes collective operations for hierarchical wide-area systems.