Disputas: Marcin Michal Krotkiewski

M.Sc. Marcin Michal Krotkiewski ved Institutt for informatikk vil forsvare sin avhandling for graden ph.d. (philosophiae doctor)

"Efficient Implementations of Numerical Models for Geological Applications on Modern Computer Architectures"

Tid og sted for prøveforelesning

29. nov. 2010 10:15 (Lille auditorium, Informatikkbygningen, Gaustadalléen 23) - A Comparison of Direct and Iterative Solvers for 2 and 3 Dimensional Simulations on High Performance Computers

Bedømmelseskomité

  • Professor Jason Phipps-Morgan, Cornell University.

  • Professor Petter E. Bjørstad, Department of Informatics, University of Bergen.

  • Senior Researcher Galen Gisler, Physics of Geological Processes, University of Oslo.

Leder av disputas

Dag Langmyhr

Veileder

  • Daniel W. Schmid
  • Yuri Podladchikov
  • Knut-Andreas Lie
  • Knut Martin Mørken

For mer informasjon

Computers have long been used by scientists to study natural phenomena through simulations. Recent advancements in parallel super-computing architectures offer the possibility to model many scientific and engineering problems with unprecedented accuracy, and in practical time. This work demonstrates how modern high-performance computing platforms (multi-processor workstations, massively parallel clusters with thousands of CPUs, and high-end graphics cards) can be used to perform efficient and accurate numerical simulations of geologically relevant physical processes.

Computer simulations require that the model is discretized using a computational mesh with a finite number of points. The geometrical complexity and the three-dimensional nature of geological systems necessitate models with high resolutions, i.e., large number of mesh points. Often, meshes with hundreds of millions to billions of nodes are required. The amount of computational work that needs to be performed grows with the number of nodes. Thus, performing the computations on a single processor would take a very long time. Moreover, in many cases the size of the studied problems can by far exceed the memory capacity of ordinary desktop computers.

Parallel computing is the answer to these difficulties: by dividing the data and the work to be done, many processors can potentially compute many times faster than a single CPU. This thesis presents several efficient implementations of numerical models for a number of modern parallel platforms. The major concern is the code performance, i.e., optimal utilization of the hardware. Several new optimization techniques are presented that speed up the codes on modern multi- and many-core architectures. Optimality of the implementations is demonstrated and parallel scalability is shown on clusters with thousands of CPUs, on workstations with tens of CPUs, and on high-end Graphics Processing Units. Finally, usefulness of the developed models is shown in several geologically relevant simulations.

Kontaktperson

For mer informasjon, kontakt Else Marie Lingaas.

Publisert 25. feb. 2011 09:58 - Sist endret 25. mars 2014 10:51