Neural Network Based Analysis of High Performance Computer Resources Utilization
Tajti Tibor
<>
Debreceni Egyetem IK
Gál Zoltán Dr.
<>
Debreceni Egyetem TEK
The analysis using artificial neural networks is very useful for problems which have rich
datasets while not having any known algorithm for their solution and the available dataset is not
complete, because of containing missing or invalid data. The available data has many interrelations,
which in the proper form can be understood as parameters.
The virtualization of rural supercomputer nodes of the Hungarian higher education and
academic system requires the coordination of computational and storage resource sharing of the three
different HPC machines. In this task the system's resources used by different research computer
programs performing computation on different target tasks are analyzed. For the first approach from
the three different HPC architectures we analyze the MPP in Debrecen, but the method that will be
developed, will also be applied for SMP and blade HPC systems.
For the precise formulation and solution of the problem the standard questions of neural
networks will be answered: the neuron transfer function, the neural network topology and its
correlation to the CPU's physical topology, the neuron types and the number of them in the different
layers, as well as the training algorithms and their parameters.
In this presentation we plan to provide an overview of the applied research work that is going
on in the TÁMOP 4.2.2.C-11/1/KONV-2012-0010 project at the University of Debrecen in relation to
the development of the HPC system for more efficient sharing of resources.