This survey provides a background information, state-of-the-art, and analysis of petascale interconnection networks. Based on this, an extrapolation of technological changes (topology, routing and flow-control algorithms) and physical constraints (port counts, number of communication nodes and communication speed) is made to predict the optimal exascale interconnection network. Such an approach leads to the conclusion that a high radix topology with optical connections for longer links are optimal interconnect for most application domains.
COBISS.SI-ID: 29774119
The current era of supercomputing is referred to as the Petascale (10^15) era. The next big HPC challenge is to break the Exascale (10^18) barrier. Due to technological limitations, the achievement of this goal will require a substantial shift towards hardware/software co-design, e.g. a computational dataflow in custom-designed hardware accelerators. This, in turns, requires a performance metric that (a) is highly meaningful and (b) to a high extent reflects the real, usable performance even in unconventional computational approaches. Our viewpoint is that one needs a multidimensional metric that covers additional parameters to the number of floating point operations (FLOPS), such as as, performance per watt, performance per cubic meter, or performance per monetary unit. This publication significantly contributes to the reputation of our research programme, for it results from cooperation with a group of internationally eminent experts known in the field for classification of parallel computers, for early introduction of microprocessors and GaAs technology, and for commercialization of high performance dataflow computers.
COBISS.SI-ID: 26715687
We have proposed a new approach for online construction of classifiers based on generative models for probability density estimation. The model was derived from the online Kernel Density Estimation approach, oKDE, which we have published in Pattern Recognition journal in 2011. The new approach incrementally builds classifiers from streams of data by means of KDE. In contrast to oKDE, the new approach applies a discriminative cost function in the model simplification. Consequently, it results in models that are more compact and have a higher discriminative power. It yields discriminative performance comparable to that of batch approaches, but with smaller computational complexity, which allows fast construction of generative models with high discriminative power. The proposed approaches are especially suitable for robotic systems proposed in this project, since they allow online construction of models from streams of data at low computational complexity.
COBISS.SI-ID: 9907284
Through detailed analysis of accesses to different memory levels (L1-L3 caches), we demonstrated and explained super-linear speedup in dealing with the unsteady thermo-fluid problem on a 4-processor system with a total of 40 cores. During the execution of the original simulation program, we counted accesses to different levels of cache. Analysis of the results showed that the super-linear speedup results from the combining of several L3 caches. The main message of the paper is that the execution time of the program strongly depends on the architecture of the computing system and not only on the power of the processors.
COBISS.SI-ID: 27339815
In the book "Parallel Scientific Computing", we present the background and motivation for the development of solution methodologies for partial differential equations. The book is concentrated on the synergy between computer science and numerical analysis. It is written to provide a firm understanding of the described approaches to computer scientists, engineers or other experts who have to solve real problems. Different solution approaches are described in detail, with a description of the required algorithms and methods that are needed for the design of efficient computer programs. Most of the details are demonstrated on solutions of practical problems, from basic to more complicated ones. This book is a useful tool for readers interested in the implementation of complex computer simulations in various application domains.
COBISS.SI-ID: 28468007