High-performance computing in geoscience [Elektronische Ressource] : data preprocessing by domain decomposition and load balancing / vorgelegt von Dany Kemmler

Dipl.-Inform. Dany Kemmler, School Of Geology, University Of Tuebingen, Email: <[e-mail]>

Sujets

• Informatik
• Mathematics

Informations

• Publié par: eberhard_karls_universitat_tubingen
• Publié le: 01 janvier 2009
• Nombre de lectures: 13
• Langue: English
• Poids de l'ouvrage: 11 Mo

Extrait

High-Performance Computing in
GeoScience
Data Preprocessing by Domain Decomposition
and Load Balancing
Dissertation
zur Erlangung des Grades eines Doktors der Naturwissenschaften
der Geowissenschaftlichen Fakultät
der Eberhard Karls Universität Tübingen
vorgelegt von
Dany Kemmler
2009Tag der mündlichen Prüfung: 16.03.2009
Dekan: Prof. Dr. Peter Grathwohl
1. Berichterstatter: Prof. Dr. Paul D. Bons, Tübingen
2. Senior Lecturer Dr. Sven Helmer, LondonIn memory of my beloved grandma, who passed away
too soon. Thank you to Matthias, for his love,
care and never-ending support, and also to my
family and all the countless good souls everywhere
who believed in me, encouraged me and helped me,
so that this thesis could become reality one day ...Foreword and Acknowledgements
When my supervisors and the administrative people at the School of Geosciences here in Tuebin-
gen are asked what part of my thesis would be the most important one, they definitely will reply:
“Everything from page 1 on.” In my opinion, this is just half the truth! They might be right in
regard to the pure “facts” and the “hard work” that is always hidden behind such a big and long
Ph.D. project.
But, to be honest, in my eyes this short chapter here is really the most important part. Such
a thesis could never succeed without the help of countless people in the background who give
support, offer their help, have an open ear for all kinds of problems and who do not hesitate to give
encouragement, especially at those times when everything looks quite hopeless and the seemingly
endless amount of work appears to exceed the capacity of one human being.
It is impossible to name them all; each is a single jigsaw piece in the whole project called “Ph.D.
thesis.” It doesn’t matter in what way the support was offered. Everybody is equally important.
Therefore, thank you to you all!
Nevertheless, I explicitly would like to thank some of these helpful souls here in personal by name.
The order is not important; I have to start somewhere ...
First of all, there is my “better half,” Matthias, who never stopped believing in me, even at times
when I was completely convinced that working towards a Ph.D. was the most stupid idea I had ever
had in my whole life. His loving support and continuous encouragement through this seemingly
endless time of writing my dissertation is unbelievable. Thank you, honey! Without you this work
would never have been finished. Thank you for believing in me and being there for me all along!
As you can see, this thesis is written in English, although I’m German. I have no English-speaking
roots nor am I a linguistic talent. But when you peruse this text it reads smoothly and looks
perfectly fine. Well, this is the hard work of Ginny Dittrich! I can’t thank her enough for her
endless patience in correcting my “colloquial wish-wash” and somehow converting the whole into
the nice and readable English thesis that you now hold in your hand. She was my “Language
Dschinn” and I can count myself more than lucky that she agreed to help me with my English!
Although it is “their job,” I definitely would like to thank both my Ph.D. supervisors, Paul Bons
in Tuebingen and Sven Helmer in London. There was a time when it looked as if I could finally
“bury” the prospect of ever finishing my Ph.D. thesis. With their help and support it was possible
to finish this work, despite administrative problems and unlucky circumstances which cropped up.
They have been very helpful, have given good advice, have been very fair and had very good ideas.
I experienced a great deal of support from both of them. They were not required to help me, but
offered me a hand and made my dream of a Ph.D. come true in the end.
Another person whom I can’t thank enough is Wolfgang Frisch. He always had an open ear.
Whenever I thought there was no way out of mountain ranges of problems of all kinds, he always
5had a solution or an idea and a word of comfort. Without his endless support and encouragement I
would never have been able to finish this thesis. And for me a little dream came true when it turned
out that I would be the very last Ph.D. candidate he would be testing before finally retiring from a
long academic life. I hope I didn’t disappoint him in the end ...
One person unfortunately is no longer able to celebrate the termination of my thesis with me. This
is my beloved grandma who passed away way too soon. She was the very first one to encourage
me to step forward and start such a “big project” and to work towards my Ph.D. She was always on
my side even when countless others thought I was completely out of my mind. She was so proud
and was looking forward to the day when I finally would step up to her door and ring her bell with
my Ph.D. certificate in my hand. Wherever she is now, this thesis is dedicated to her and I hope
she is proud and very happy. Thank you grandma; I miss you!
In addition to the people mentioned above, there are also Steffen and Hadmut who helped me
whenever the technical equipment refused to work as I wanted it to. They have been my “computer
emergency squad” and I always could count on them. Then there is of course my family, who first
seemed a little bit “surprised” about my “stupid” decision to go for my Ph.D., but never gave
up hope and believed in me. And last but not least I also am grateful to the fantastic group of
people around Paul Bons and Wolfgang Frisch who have always been very nice to me, have been
supportive whenever I needed help or just someone to talk to, and who accepted me as if I had
always been a normal and full group member. They provided a very pleasant and friendly working
environment and I enjoyed being there.
Most likely I have forgotten a million other helpful souls. In case someone is missing his or her
name here, I’m very sorry, but to all of you a big thank you and my gratefulness for everything you
all did for me so that I finally could hand in this piece of work that represents my Ph.D. thesis.Höchstleistungsrechnen in den Geowissenschaften –
Datenvorverarbeitung mittels Gebietszerlegung
und Lastbalancierung
Computer sind heute ein fester Bestandteil in unserem täglichen Leben und haben zu einem ex-
tremen Fortschritt in den letzten Jahrzehnten beigetragen. In Forschung und Wissenschaft hat
der Einsatz von Computersimulationen die Beantwortung vieler wichtiger und spannender Fragen
erst ermöglicht. Ohne solche Simulationen können bestimmte, komplexe Sachverhalte gar nicht
angegangen werden, weil z.B. der physische Zugang fehlt oder die betrachteten Zeiträume zu lang
sind. Die Untersuchung von Prozessen in vielen 100 km Tiefe in unserer Erde, das Verhalten von
radioaktivem Müll, der in einem in Bentonit umschlossenen Metallcontainer eingelagert wird, und
den dann stattfindenden Prozessen über viele 1000 Jahre oder präzise Wettervorhersagen auf dem
Niveau, wie wir es heute gewohnt sind, gehören zu diesen Sachverhalten.
Aber auch trotz des enormen Fortschritts im Sektor der Computertechnik ist es heute immer noch
nicht möglich mit Standardcomputern sehr rechenintensive Probleme zu lösen. Hier ist der Einsatz
von Spezialhardware von Nöten, die mehrere 10 Mio. Euro kostet. Erst durch die Nutzung von
diesen Höchstleistungsrechnern sind z.B. die heutigen Wettervorhersagen überhaupt erst möglich.
Aber auch wichtige Fragestellungen in vielen anderen Gebieten der Wissenschaft können nur mit
Hilfe solcher “Number Cruncher” angegangen werden.
Für Computersimulationen erstellt man allgemein immer erst ein Modell oder Abbild der Realität,
mit dem dann die Antworten auf bestimmte Fragen berechnet werden sollen.
Bei der Entwicklung von Software, die innerhalb solcher Modelle die vorhandenen Daten ver-
arbeitet und Berechnungen vornimmt, ist es wie mit der Entwicklung eines Formel-1 Wagens:
Letztlich handelt es sich, aus der Sicht eines Laien betrachtet, immer noch um ein Auto; tat-
sächlich ist aber jeder einzelne Formel-1 Wagen eine genau angepasste Einzelanfertigung, an der
viele, verschiedene Ingenieure mitentwickelt haben. Genauso verhält es sich mit Software, die auf
Höchstleistungsrechnern einmal laufen soll. Für den Laien ist es nichts anderes als ein bisschen
Software in C oder Fortran, doch tatsächlich handelt es sich bei diesen Programmen um exakt
abgestimmte Spezialsoftware, die genau auf die jeweils unterliegende Hardware angepasst sein
muss. Viele einzelne Punkte im Parallelisierungsprozess sind wichtig. Aber nur die Gesamtheit
aller notwendigen Einzelschritte ergibt letzten Endes ein lauffähiges, genau auf die Maschine aus-
gerichtetes Höchstleistungscomputerprogramm.
Typischerweise zeichnen sich die Modelle, die mit den Höchstleistungsrechnern berechnet und
simuliert werden sollen dadurch aus, dass sie eine enorme Menge an Eingabedaten bewältigen
müssen. Eine gezielte Vorverarbeitung dieser Datensets mit oft mehreren 10er Millionen Eingabe-
daten, ist ein unabdingbarer Schritt, um im späteren Verarbeitungsprozess, die eigentliche Problem-
oder Modellparallelisierung erst angehen und durchführen zu können. Nur wenn die Eingabedaten
optimal auf- und vorbereitet sind, kann die unterliegende Spezialhardware sinnvoll für die spätere
Modellberechnung genutzt werden.
Trotz der hochwerti