We present a parallel, medical application for the analysis of dynamic positron emission tomography (PET) images together with a practical performance model. The parallel application improves the diagnosis for a patient (e. g. in epilepsy surgery) because it enables the fast computation of parametric images on a pixel level in contrast to the traditionally used region of interest (ROI) approach. We derive a simple performance model from the application context and demonstrate the accuracy of the model to predict the runtime of the application on a NOW. The model is used to determine an optimal value for the length of the messages with regard to the per message overhead and the load imbalance.

This paper deals with currently used algorithms for the reconstruction of functional images which run up to 60 hours or more on a single workstation and deal with hundreds of megabyte of data. A parallel implementation with high efficiency and almost linear speedup of a sophisticated iterative algorithm is given and its applicability to other reconstruction methods is shown. Whereas running this application on a high performance parallel computer is straightforward, there are more issues under production conditions as they are enforced by daily routine in a clinic. We adress the topic of fault tolerant parallelizing and batch queuing of programs which are typically written in a high level language like IDL or MATLAB and show how load balancing can preserve the ownership of workstations in a network of workstations (NOW) which is used for distributed computing during office hours.

We present the first parallel medical application for the analysis of dynamic positron emission tomography (PET) images together with a practical performance model. The parallel application may improve the diagnosis for a patient (e. g. in epilepsy surgery) because it enables the fast computation of parametric images on a pixed level as opposed to the traditionally used region of interest (ROI) approach which is applied to determine an average parametric value for a particular anatomic region of the brain. We derive the performance model from the application context and show its relation to abstract machine models. We demonstrate the accuracy of the model to predict the runtime of the application on a network of workstations and use it to determine an optimal value in the message frequency-size relationship.

The ARB (from Latin arbor, tree) project was initiated almost 10 years ago. The ARB program package comprises a variety of directly interacting software tools for sequence database maintenance and analysis which are controlled by a common graphical user interface. Although it was initially designed for ribosomal RNA data, it can be used for any nucleic and amino acid sequence data as well. A central database contains processed (aligned) primary structure data. Any additional descriptive data can be stored in database fields assigned to the individual sequences or linked via local or worldwide networks. A phylogenetic tree visualized in the main window can be used for data access and visualization. The package comprises additional tools for data import and export, sequence alignment, primary and secondary structure editing, profile and filter calculation, phylogenetic analyses, specific hybridization probe design and evaluation and other components for data analysis. Currently, the package is used by numerous working groups worldwide.

We describe an accurate method for the automatic parallel generation of oligonucleotide probe sets for DNA microarrays. This approach includes a component for high-performance specificity evaluation of designed probes in large data sets. The three main algorithmic components of the method, namely probe preselection, hybridization prediction and probe selection are explained in detail. We introduce new combinatorial techniques for the efficient selection of probe sets of high differentiation capability even from sequence databases of conserved homologous genes. These techniques include the automatic generation of group specific probes as well as the design of excluding probes. A basic prototype has been implemented including a shared memory parallelization. Test runs have been performed on a multiprocessor personal computer with subsets of a small subunit ribosomal ribonucleic acid database, containing very conserved sequence data. The applicability of our program is pointed out by designing a set of oligonucleotide probes that shall allow a comprehensive parallel identification and differentiation of several groups of extremophilic prokaryotes by DNA microarray. The probe set is accessible via the Internet. On applying the parallel version on a dual processor system an efficiency of 80% was achieved. Copyright © 2004 John Wiley & Sons, Ltd.

Ziel dieser Arbeit ist die Evaluierung effizienter Scheduling- Algorithmen für den praktischen Einsatz in der medizinischen Bildverarbeitung, am Beispiel der Spektralanalyse dynamischer Aufnahmen der Positronen-Emis-sions-Tomographie (PET). Die Spektralanalyse, die im Gegensatz zu klassischen Kompartmentmodellen zur bildpunktweisen Berechnung physiologischer Parameter ohne a priori Modellannahmen eingesetzt wird, stellt ein wichtiges Werkzeug zur Analyse dynamischer PET-Aufnahmen dar. Die parallele Berechnung parametrischer Bilder ermöglicht den Einsatz modernster Analysemethoden auch für klinische Datensätze, beispielsweise im Rahmen der Epilepsiechirurgie. Die Spektralanalyse repräsentiert durch die hohe inhärente Parallelität eine charakteristische Anwendung der funktionellen Bildgebung. Durch die verteilte Berechnung der physiologischen Parameter auf einem Network of Workstations (NOW), lässt sich die Gesamtlaufzeit der Anwendung auf ca. 3 Minuten reduzieren. Wir betrachten fünf Scheduling-Strategien, die auf den Methoden theoretischer Arbeiten über parallelisierende Compiler für speichergekoppelte Parallelrechner aufbauen.

Online visualization and computational steering of parallel scientific applications has been widely recognized as the key to better insight and understanding of the observed simulation. From the parallel program developer's point of view further problems arise and need to be solved. The behavior and performance of parallel programs does not only depend on the input data but also on inter-process communication. To reflect this fact we propose a novel combination of online visualization, computational steering of parallel high performance computing applications and controlled deterministic execution. Both the visualization and the classical part of steering is based on the VIPER tool. For the control of the communication we rely on a tool called codex, which was developed to test and control communication by the use of control patterns. Finally, VIPER and codex form an environment for tuning, steering and testing based on VIPER's extended programming model.

This paper presents the design and implementation of NSR- the Node Status Reporter. The NSR provides a standard mechanism for measurement and access to status information in clusters of heterogeneous workstations. It can be used by any application that relies on static and dynamic information about this execution environment. A key feature of NSR is its flexibility with respect to the measurement requirements of various applications. Configurability aims at reducing the measurement overhead and the influence on the observed cluster.

This paper deals with the parallel implementation of reconstruction algorithms for functional imaging on a network of workstations (NOW). Algorithms which provide the best image quality are not used in clinical routine, because they have a runtime of up to 60 hours with real clinical data sets of several hundred megabytes. After giving an overview of currently used image reconstruction algorithms, we describe a general parallel implementation of these algorithms with almost linear speedup and high efficiency which cuts down the runtime to a feasible limit. The high load which is caused by the parallel application conflicts with the predominantly interactive usage of clinical workstations, therefore we address load balancing with an application oriented, adaptive mechanism in order to preserve the ownership of workstations. Furthermore we explain how the integration of MATLAB and IDL based applications with a conventional distributed queuing system (DQS) can be achieved and why this significantly improves usage in clinical routine.

The on-line monitoring interface specification OMIS pro vides means for developing more powerful interoperable and portable tool environments for parallel and distributed systems. It specifies the interaction between any tool and a monitoring system that is responsible for observing and manipulating the programs' execution. This well- defined interface makes it possible to concurrently use several tools of possibly different developers with the same program run and to port tools onto various target architec tures and software environments. As a starting point, the research group at LRR-TUM is designing an OMIS com pliant monitoring system OCM for Parallel Virtual Machines PVM to run on workstation clusters. Tool de velopers can use this implementation to attach their own on-line tools to the system.

This article covers research at Technische Universitat Munchen on distributed and parallel architectures and applications. First, an overview on the parallel processing research organization is given. The second main topic covers an integrated hierarchical programming environment TOPSYS for parallel and distributed systems developed as part of the research grant.

Numerical simulation requires very high performance computer architectures. This can only be achieved by using various forms of parallelism. The second section of this paper gives an introduction to the state-of-the-art in parallel computer architectures and their programming. Today, some open problems remain for the user of such architectures. Those problems and possible solutions are dealt with in the third section. Problems of communication, portability, and flexibility are addressed.

The here presented work shows an efficient parallel implementation of a 3-D Navier-Stokes solver. The code is easily transferable. The numerics remained all vectorized, enabling a vector as well as a parallel run on multiple computer systems. The implicit solver works blockwise. This technique saves memory in case of a sequential run and enables high speed-ups in case of a parallel run. Parallelization remains coarse grain for any number of nodes that are introduced. INTRODUCTION Present research and development in aerospace science disclose the strong need for more computing power and not least memory. Traditional super vector computers more and more reach physical limitations and are increasingly expensive. A different approach to obtaining more computing power and storage has been the development of multiprocessor systems with powerfull RISC processors. This technology promises scaleable computing power and memory at moderate prices. This paper deals with the parallelization of an industrial 3-D Navier-Stokes solver on multiprocessors of different architectures and programing models. With the development of general message passing calls which are realized on the particular multiprocessor through the available communication model, the code is easily transferable. Thus the code is as independent as possible from the machine architecture, which is important if one considers the dynamic market situations in that field. All loops and the numerics remained fully vectorizeable, which guarantees a good performance on vector Transactions on Information and Communications Technologies vol 3, © 1993 WIT Press, www.witpress.com, ISSN 1743-3517

at M¨ unchen