Effiziente Parallelverarbeitung mit Nachrichtenaustausch auf Metacomputern

Since the beginning, mathematical description and analysis of physical processes has been a core area for all disciplines of natural science. Over the last several decades, numerical analysis with the goal of simulating processes has increased in importance due to the emergence of digital computers. The requirements of increasingly precise simulation results while simultaneously facing increasingly complex models quickly proved too much of a challenge for the capacities offered by simple Von-Neumann computers. One possibility of boosting a system's computing power to match these increasing requirements is by using several processes for the parallel processing of a given problem, where many various architectures of corresponding parallel computers have developed over time. Nevertheless, even these kinds of systems again quickly reach their capacities in the face of problems of an increasing size. These limits can lay far below the immense computing power requirements needed to solve large-scale problems, such as, for example, reliable climate simulation. As a result, a further logical step would be to link several such parallel systems into an even higher-order unit, which we classically term a metacomputer. The nature of a metacomputer features hierarchical and often heterogeneous structures which must be taken into consideration when creating an efficient mapping of parallel algorithms using this form of architecture. Within such systems, a message exchange is normally used as a parallelization paradigm, in which the parallel processors send the intermediate computation results explicitly through correspondingly hierarchically-structured networks. In order to realize this message exchange at the application level, the so-called Message Passing Interface (MPI) normally is resorted to, which represents a programming interface to the underlying communication libraries. Two key aspects must be considered in order to design this kind of message passing-based parallel processing on a metacomputer in the most efficient way possible. First, the underlying communication library must be able to fully utilize the various technologies within the system in the most optimal manner. Second, the parallel algorithm must be adapted to the hierarchical and heterogeneous system by using smart mapping, allowing the algorithm to use the available resources in the most efficient way possible. This dissertation will discuss both of these requirements in detail, while taking into consideration various other aspects. In doing so, the focus of discussion will be on the realization of a metacomputer-enabled MPI library, which the author has played a key role in co-developing over the past several years at the Department of Operating Systems at the RWTH Aachen University