The research of the PVS-group has a focus on high performance computing and storage. We are interested in high performance hardware infrastructures, operating system and middleware components, and application programs that make use of this computer architecture.

With scientific computations we observe a dramatic increase in the amount of data that is computed. Even with modern high performance computers the storage capabilities often form the bottleneck in computing more detailed results. We focus our research and development interests in this field: the high performance input/output concepts and systems. Research is conducted with parallel file systems, cluster environments, and parallel applications that have a high I/O demand.

Research is focussed in the field of high performance input/output systems and concepts. The following list covers the major research interests:

• Performance Evaluation of High Performance I/O Systems
  High performance I/O systems exhibit certain characteristics that are crucial for the overall performance: Scalability with respect to increasing numbers of clients and servers is one such aspect as well an efficient handling of files of various sizes. We compare different systems with respect to their performance characteristics and also take into account other aspects like e.g. fault tolerance, configurability etc.

• Visualization of Performance Data with High Performance I/O
  For the user as well as the developer of an I/O system it is essential to get some insight into the behaviour of the whole system during operation. We base our work on available trace-based performance tools and enhance them in a way that visualizes the internal activities of the I/O middleware. The crucial point is to relate user-level operations to these low-level activities in order to be able discuss performance influences of changes here and there.

• Configuration of Cluster Environments for Parallel I/O
  Clusters have so far mainly been used as computational resources. However, as many of them have integrated hard disks we can also use them as a configurable I/O resources. Similar concepts have to be applied as with compute nodes: How many of them do I use for what time? What overall I/O performance can I expect? How persistent will the files be stored? How will node and disk failures be handled?

• The Influence of I/O Patterns onto Performance
  When files get distributed onto different disks with parallel I/O the question arises which low level data file is hit when a certain byte of the parallel file is accessed. Parallel file systems determine these locations by means of distribution functions: they map from a file position to a position in a data file. Overall performance is influenced by the appropriate selection of such a function. The function is chosen with respect to the file's I/O access pattern, i.e. the spatial and temporal distribution of access to individual bytes.

• Deployment of Parallel I/O in Sequential and Parallel Programs
  There is not yet so much knowledge available about the deployment of parallel I/O concepts in parallel programs. Few MPI-based programs do already make efficient use of MPI-IO. As the number of available I/O systems is increasing we will see more such programs in future. As with parallel programming it will take quite some time to learn how to use parallel I/O efficiently and to adapt sequential I/O programs to parallel I/O programs.

• Organic Computing
  A general problem of modern distributed computing systems is to master their increasing complexity. Organic Computing provides a specific concept of how to cope with complexity. This biologically inspired concept is based on goal-oriented self-organization and adaptation. Many biological systems exhibit properties of self-organisation, like a high degree of scalability, robustness and self-optimization, which are desirable in large scale distributed computing systems. Since self-organisation occurs only in systems of many interacting units, distributed computing systems appear to be perfectly suited for a realisation of organic computing in an artificial system. Nevertheless, there are still difficulties resulting from the digital data format that is an intrinsic part of the current Information Technology, but no characteristic of the biological paradigm.

Thomas Ludwig is active in the field of parallel and high performance computing since the mid 80s. Over the years he worked with many different generations of parallel computers mostly deploying distributed memory as a architectural characteristic. He developed environments for the on-line observation and manipulation of parallel programs: debugging, performance analysis, program flow visualization, and load balancing. In later years the research concentrated on a systematic design of internal interfaces from tools to monitoring systems that provide means for the necessary program interaction. Adapting programs to high performance computers was also always of importance: he conducted research and development for program parallelization for computer tomography in medical science and for the computation of phylogenetic trees in bioinformatics.