The brochure of the John von Neumann Institute for Computing is available in English and in German. It can be ordered at the NIC secretariat (nic@fz-juelich.de).

Introduction	Super- computing	Astro- physics	Elementary Particles	Many Particles	Polymers	Chemistry	Environment	Other Fields

    Supercomputing

Text "Supercomputing"	Wolfgang Gürich, NIC-ZAM
Cray Complex in Jülich	NIC-ZAM
Schematic Representation of the JuNet Subnetworks and their Connections	Dieter Conrads, NIC-ZAM
APE100 Installation at DESY Zeuthen	Ulrich Gensch, Peter Wegner, NIC-DESY Zeuthen
Standard PCI Bus Card	Ulrich Gensch, Peter Wegner, NIC-DESY Zeuthen
Response Time Monitor	Wolfgang Gürich, NIC-ZAM
System Supervision with TREND	Bernd Mohr, NIC-ZAM
Metacomputing in the Gigabit Testbed West	Thomas Eickermann, NIC-ZAM
UNICORE: Uniform Interface to Computing Resources	Dietmar Erwin, NIC-ZAM
Automatic Performance Analysis of Parallel Programs with KOJAK	Michael Gerndt, NIC-ZAM
Development of Algorithms for DYNA3D	Johannes Grotendorst, NIC-ZAM

Text "Supercomputing"

Scientific computing has already played an important role for many years on an international level in the large research institutions. The Research Centre Jülich entered the aera of supercomputing in 1983 with the installation of the CRAY X-MP computer, the first vector computer of this power class in Europe. Scientific computing has since been consistently further developed towards supercomputing, simulation and modelling by exploring new algorithms and methods, extending the range of applications as well as expanding innovative computer resources.

The computer resources currently available for NIC in Jülich are based on the supercomputer complex realized in 1997; components are three vector computers and two massively parallel computers connected by a high-speed network.

For applications in the field of theoretical high energy physics NIC has parallel computers with SIMD architecture installed at DESY-Zeuthen. This computer architecture has proved particularly efficient especially in the simulation of quantum chromodynamics.

In order to offer NIC users the optimum benefits from the computer resources available, research and development for incorporating the supercomputers into computer operation, for connection to high-speed networks and for programming support are indispensable. These activities must be closely related to the necessary further development of hardware and software.

In addition to systems for availability monitoring and control, tools are also used for evaluating the quality of use (RTM, TREND). The possibility of distributed computing on different computer architectures connected via broadband communication networks is being tested (metacomputing). The development of uniform, platform-independent computer access for users creates the technical prerequisites for the cooperating of supercomputing centres (UNICORE).

NIC users will find direct support from project-accompanying consulting services, courses and exchange of experience at user workshops. The approach to programming parallel programs is facilitated by the development of methods for the automatic execution of analysis tools (KOJAK).

New parallel algorithms are gaining entry into program libraries and also into the porting of program packages on massively parallel computers, and e.g. for applications from structural mechanics they provide new orders of magnitude in the simulation of more complex systems (DYNA3D).

(Wolfgang Gürich, NIC-ZAM)

Cray Complex in Jülich

The Cray supercomputer complex in Jülich is composed of the massively parallel systems CRAY T3E-600 with 512 processors (300 GFLOPS) and CRAY T3E-1200 with 512 processors (530 GFLOPS) as well as the vector computers CRAY T90 (20 GFLOPS) with 10 processors and CRAY J90 (3.2 GFLOPS) with 16 processors. Another CRAY J90 system with 12 processors is used as a file server. All Cray systems are connected by a fast network (GigaRing with 1.6 GByte/s).

Schematic Representation of the JuNet Subnetworks and their Connections

ZAM's task is to provide efficient communication systems according to the respective state of the art. The local network of the Research Centre, JuNet, consists of the established Ethernet/FDDI part, complemented by a new component with the installation of an ATM network which was commenced in 1995. JuNet connects the internal institute networks with each other, with central and decentralized server capacities at the Research Centre and with the outside world.

Apart from ISDN, which is primarily used for access to JuNet from the private sphere, the Research Centre's worldwide network integration is effected via the broadband science network (B-WiN) of the DFN Association (Association for the Promotion of the German Research Network), currently with a 34 Mbps link. As project leader of the Gigabit Testbed West, one of two Gigabit Testbeds in Germany, the Research Centre participates in the preparations for G-WiN (Gigabit-WiN) as the successor to B-WiN.

In addition to existing services, the introduction of the ATM technology into JuNet will create the prerequisites for high-resolution visualizations, for the use of modern broadband multimedia network services and for metacomputing, both within the Research Centre and - initially project-oriented within the framework of the Gigabit project - to external partners. Prior to the introduction of innovative communication techniques, ZAM is carrying out pilot implementations of new network techniques and beta tests of new equipment generations.

(Dieter Conrads, NIC-ZAM)

APE100 Installation at DESY Zeuthen

The Centre for Parallel Computing at DESY Zeuthen has been established for processing particularly compute-intensive problems in the theory of elementary particles. Since mid-1994 "Quadrics" parallel computers from Alenia Spazio have been installed at Zeuthen. These computers are based on the APE100 development by Italian elementary particle physicists from the Istituto Nazionale di Fisica Nucleare (INFN) and offer a particularly good price/performance ratio for simulations on lattices. In contrast to conventional supercomputers they work according to the SIMD principle (single instruction, multiple data) with all processors operating completely synchronously. They represent, in particular, a highly efficient and low-cost solution for computations in the field of lattice field theories permitting numerical access to the basic properties of elementary particles and their interactions.

The current installation at DESY consists of three computers with a total of 768 processors and a peak performance of 45 GFLOPS. The computers have been running since early 1994 almost without any problems. On an annual average, a utilization of > 90 % was achieved. The ratio of achieved to maximum possible performance was in the range of 30 to 70 per cent depending on the application.

(Ulrich Gensch, Peter Wegner, NIC-DESY Zeuthen)

Standard PCI Bus Card

Based on the present APE100 architecture, the successor model has been designed at INFN, representing a step forward on the road to the 1 TFLOPS computer. In order to achieve this goal, DESY contributes to the APEmille development project in the field of communication hardware and software development. A card based on the standard PCI bus was developed which permits data transfers of up to 132 MBytes/s between two PCs via serial communication interfaces. These cards connect all so-called host PCs controlling the massively parallel APEmille computer. Moreover, this card can be used for networking conventional PCs or workstations in the high speed range. The APEmille architecture also eliminates some restrictions of the APE100 family by providing 64-bit floating point arithmetics, local addressing and faster communication interfaces which lead to broader application possibilities.

The cooperation between INFN and DESY will be further intensified in future in order to design and develop joint innovative parallel computers for elementary particle physics beyond the APEmille project.

(Ulrich Gensch, Peter Wegner, NIC-DESY Zeuthen)

Response Time Monitor

The response time behaviour is an important factor for evaluating the performance of interactive systems. In order to measure this behaviour and to detect bottleneck situations at an early stage, a response time monitor (RTM) was developed and installed for supervising the central interactive computer systems in Jülich. An RTM agent simulates an interactive session of a "standard user" by a sequence of commands, program requests and "pauses for reflection". The measured response times and other parameters (e.g. system utilization, number of active users) are transmitted to a WWW server which makes the data available as an HTML document for the Internet.

(Wolfgang Gürich, NIC-ZAM)

System Supervision with TREND

TREND (Torus REsources and Node Display) allows the supervision and observation of a CRAY T3E computer system. TREND is a distributed client/server system. It consists of a TREND data acquisition demon running on the system to be supervised, which captures all necessary data once a minute and transfers them to a conventional WWW server. In this way, the data can be accessed from anywhere in the world. The server distributes the data to the TREND display clients running on the users' local workstations. The display comprises information on all running, blocked and waiting user programs, on configuration data of the overall system and of single processing elements, and on utilization. Furthermore, general system information is provided.

The development and marketing of TREND is supported by SGI/Cray Research. TREND is currently being used in computer centres all over the world.

(Bernd Mohr, NIC-ZAM)

Metacomputing in the Gigabit Testbed West

The computing power of even the most up-to-date supercomputers is not always sufficient for the treatment of sophisticated numerical problems. The increasing stability and bandwidth of modern wide-area networks (WANs) allows high-performance computers at various sites to be connected to a virtual overall system so as to build up a so-called metacomputer.

Within the framework of a BMBF-funded project, the Gigabit Testbed West, parallel applications for use in metacomputing are being investigated. The project involves five German research institutions, three universities and two industrial companies. A powerful ATM link (currently 2.4 Gbit/s) is available between NIC in Jülich and GMD in St. Augustin.

So-called "coupled-field" algorithms distributing two or more loosely coupled position- and time-dependent variables to the computers are particularly suited for metacomputing since the demand for communication between the parts is moderate. Such an application being tested in the Gigabit Testbed West is the simulation of pollutant dispersion in the soil (in cooperation with ICG-4), where groundwater and dissolved pollutants interact. An example of "heterogeneous metacomputing", i.e. the coupling of architecturally different computers, is the real-time evaluation of magnetic resonance (MR) studies of the human brain (in cooperation with IME). An MR tomograph is coupled here to the T3E and a powerful visualization server at GMD to map brain functions and represent them three-dimensionally.

(Thomas Eickermann, NIC-ZAM)

UNICORE: Uniform Interface to Computing Resources

UNICORE is a BMBF-funded collaborative project to develop a software infrastructure which allows users intuitive, consistent and secure access to distributed computing resources. UNICORE realizes, in particular, the following functions: generation of batch jobs via a graphical user interface independent of hardware architecture and operating system; distribution of subtasks to suitable computers and synchronization of their execution by UNICORE; secure data transfer between UNICORE systems via the Internet; platform-independent job control. UNICORE applies modern web technologies such as https, signed Java applets, and certificates for secure communication and authentication. It has been designed to work in existing production environments.

UNICORE collaborators coordinated by ZAM are universities, German and European research institutions, two software companies, and vendors of high-performance computers.

(Dietmar Erwin, NIC-ZAM)

Automatic Performance Analysis of Parallel Programs with KOJAK

The application of parallel high-performance computers for the investigation of scientific problems serves to execute complex simulations within an acceptable time. The efficient use of existing resources is a prerequisite for the rapid execution of individual programs and for optimizing the overall throughput.

The typical process of performance analysis consists in the repeated execution of program instrumentation, program execution with performance measurement, and analysis of the performance data by the user until the inefficient program parts are recognized (see diagram). All three steps require great experience in handling the analysis tools and well-founded knowledge concerning possible performance bottlenecks. The aim of developing KOJAK (Kit for Objective Judgement and Automatic Knowledge-based detection of bottlenecks) is to support users in recognizing typical bottlenecks by largely automating the execution. For this purpose, a database is being built up with detection rules for potential performance bottlenecks, and an automatic control of the analysis steps is realized.

The project is embedded in the Esprit working group APART (Automatic Performance Analysis: Resources and Tools) involving three European companies, six European and three American universities under the leadership of the Research Centre Jülich.

(Michael Gerndt, NIC-ZAM)

Development of Algorithms for DYNA3D

In the field of structural mechanics suitable program packages are provided and validated on the ZAM supercomputers for projects in research and application. In addition, work is carried out for the development of parallel algorithms. In a collaborative project with the CONDAT company, a platform-independent parallel version of the CONDAT-DYNA3D finite element program has been developed.

The parallelization concept used is based on a domain decomposition of the model and the exchange of node forces by message passing. CONDAT-DYNA3D was implemented on three different parallel platforms: a workstation cluster, a massively parallel computer architecture and a shared-memory parallel computer. In the field of nonlinear structural mechanics CONDAT-DYNA3D permits applications ranging from quasi-static materials testing up to the simulation of high-speed processes.

The picture shows the controlled plastic buckling of a steel box girder with octagonal cross-section. In the model calculation, the contacts propagate uniformly over the cross-section from the impact region through the box girder. Octagonal box girders are used, for example, in vehicles to reduce the car's kinetic energy during a collision.

(Johannes Grotendorst, NIC-ZAM)

Introduction	Super- computing	Astro- physics	Elementary Particles	Many Particles	Polymers	Chemistry	Environment	Other Fields