Keynote Talk: The Beowulf Project Evolves
The Beowulf cluster project has evolved from research and demonstration systems, to prototype development systems, and is now in the early stages of being deployed as a major high performance computing platform in industrial and commercial environments.
This talk will briefly review the history of the Beowulf project, survey the state of the project, describe how Beowulf systems are being used, and discuss the likely direction of Beowulf development in the near-term future.
From the beginning the most important element of Beowulf has been building a community of developers and users. Beowulf is based on Linux. This includes not only the software itself, but the philosophy of open-source, large-scale distributed development. The project history will be described referenced to how we succeeded and failed in utilizing contributions.
The rapid acceptance of Linux in the past 18 months as the major alternative server operating system has a analogous effect in the high performance computing market. Today almost every major U.S. computer vendor has a Beowulf cluster product or strategy. These are just the first tentative steps of major change in direction from proprietary, exotic systems to standards-based commodity systems.
Parallel particle in cell codes: A comparison of the code performance on a Cray T3E, an Onyx 2, and a Beowulf Cluster, Mark Dieckmann and Anders Ynnerman
Particle in cell (PIC) codes are frequently used to simulate plasma turbulence. They represent a plasma by computational particles; their numbers are orders of magnitude less than the numbers for a real plasma. Increasing the numbers of computational particles obviously enhances the reliability of the simulation results but it also increases the simulation time. The simulation time is such that it requires (expensive) massively parallel computer resources. Finding a cost efficient parallel computer system is thus of crucial importance. We have compared the performance of our simulation code on a Cray T3E, an Onyx 2 and a Beowulf Cluster. We have found that, for our code, the type of problem we consider determines the most suitable computer system. We discuss the parallelization method of our code and two extreme physical problems that either favor a Beowulf cluster or the Cray T3E.
Project CLUX: A 96 Node Alpha Cluster, Oskari Jääskeläinen
We discuss the process of acquiring a vendor supplied Beowulf cluster for scientific research as opposed to a do-it-yourself approach. The EU negotiated procedure and main factors influencing the selection of the system are described. Five tenders are compared from a technical and performance point of view, with a look on SMP bottlenecks and system reliability. The installed system consisting of 96 Alpha processor nodes and a fast Myrinet network is described, and the status of the project is presented together with experiences thus far.
Workstation/PC clusters in a lab setting - 1 million CPU hours in 10 years, Lennart Nilsson,
Our research is focused on the dynamics and interactions of biological macromolecules, such as protein-nucleic acid complexes. Since 1991 we have used workstations and more recently personal computers to run large scale molecular dynamics simulations of these systems. Our main application programs, in terms of CPU use, are CHARMM and Gaussian98. From the humble beginnings, with two IBM RS 6000 machines, we have used NQS to handle load balancing. Today we have 51 CPUs - 7 DEC Alphas, and 44 Intel PentiumII, 450MHZ - acting as compute servers to the NQS system. The Intel CPUs, under Linux, run parallel jobs over a switched FastEthernet. We are currently installing the next generation PC cluster based on dual 866 MHz Intel PentiumIII CPUs, and here we will also have a faster interconnect, SCI.
BProc - a new framework for Beowulf clusters, Donald Becker
Scyld has introduced a new framework and set of tools for Beowulf clusters. The Scyld Beowulf system uses 'BProc', Beowulf Process Space, as a focal point of a cluster operating system distribution intended to be widely deployed for production use. This talk will describe how BProc works, how we use it to simplify installation and administration, and the benefits it provides to developers.
Preliminary results from running the NSC benchmark suite on the PC-cluster Ingvar, Torgny Faxén
The talk will present some early experiences from porting and running portions of the NSC benchmark suite to NSC's new 32 processor Beowulf cluster Ingvar. Several of NSC's most frequently used applications are part of the suite. Comparisons with the T3E and an Origin-2000 will also be presented and discussed.
The n-D SCI Torus as a basis for truly Scalable Linux Clusters for Production Environments, Knut Omang
Scali's mission is to provide cost-effective, integrated and easy-to-manage cluster solutions. For applications with high communication demands, the interconnect is one crucial factor in the choice of solution. However, as important is the software support for optimal and easy use of the interconnect to achieve this theoretical performance. We will look closer at how the high end SCI based cluster solution from Scali scales with respect to performance as well as with respect to cost and complexity. Finally, a brief introduction to the user interface to Scali's clusters will be given.
High Performance Computing on the HPC2N Linux Cluster - Benchmarks and Applications in Computational Physics, Sven Öberg
High Performance Computing Center North (HPC2N) built a Beowulf cluster last summer to evaluate it's potential for high performance computing. The cluster hosts 8 dual pentium III processors with Wulfkit type of high performance communication. Wulfkit combines Dolphin's high-speed low latency SCI interconnect technology with Scali's SCI "tuned" MPI. I will present results from benchmarks, standard and application based. The cluster is compared with some frequently used platforms for HPC.
The X86 Architecture from a Beowulf Point of View, Peter Kjellström
The most common architecture used today for building cheap Beowulf clusters is the x86 architecture. The reason often being its price/performance advantage over other platforms. With its roots in the days of the first PC, the 8086, x86 has evolved (maintaining software compatibility) into today's Intel PentiumIII and AMD Athlon based platforms. This talk will try to give a reasonably detailed description of the platform as it is today, and at the same time look into the differences between the x86 and other, more classical high performance platforms.