Table of Contents

Software Labs

Hadoop/Hive-Benchmarks

NameDaniel Pajonzeck , Stephan Succo
AdvisorsDr. Julian Kunkel, Marc Wiedemann
Date2014-04-28

Report BibTeX Sources

Hintergründe zum Projekt: Parallelrechnerevaluation

NameJürgen Kosemund, Benjamin Stark, Daniel Vach, Tim Hauschild
AdvisorsDr. Julian Kunkel
Date2013-10-29

Report BibTeX

Clustertools

NameKim Johannes Zülsdorff
AdvisorsDr. Michael Kuhn, Dr. Julian Kunkel
Date2013-09-20

Report BibTeX

Leistungsanalyse von ZFS on Linux

NameHajo Möller
AdvisorsDr. Michael Kuhn
Date2013-03-28

Report BibTeX

PySpank: Python Interface zu SPANK

NameMarina Shvalova, Oliver Bestmann
AdvisorsDr. Julian Kunkel
Date2012-09-15

Report BibTeX

Entwicklung eines Werkzeugs zur Kommunikationsanalyse von MPI-Implementierungen

NamePaul Lindt
AdvisorsDr. Julian Kunkel
Date2012-11-22

Report BibTeX Presentation Sources

Mrut - An Interface between Munin and the PIOSimHD Tool Suite

NameAlexander Njemz
AdvisorsDr. Julian Kunkel
Date2012-11-20

Report BibTeX Presentation Project homepage

Evaluation of the deduplication file system Opendedup

NameSteffen Göttsch
AdvisorsDr. Michael Kuhn, Amandine Pignier
Date2012-11-24

Report BibTeX Sources

Fortan Parallelisierung, Schwerpunkt OpenMP

NameAnna Fuchs
AdvisorsPetra Nerge
Date2012-11-16

Report BibTeX Presentation Sources

Evaluation of Distributed File Systems

NameTim-Daniel Jacobi, Jan Lingemann
AdvisorsDr. Michael Kuhn, Dr. Julian Kunkel
Date2012-10-22

Report BibTeX Sources

Cluster Management - High Performance Cloud Computing mit MAAS, Juju und OAR

NameJohann Weging
AdvisorsDr. Michael Kuhn, Dr. Julian Kunkel
Date2012-07-10

Report BibTeX

Streets4MPI

NameJulian Fietkau, Joachim Nitschke
AdvisorsDr. Julian Kunkel
Date2012-05-12

Report BibTeX Presentation Project homepage

Simulation von MPI-Collectives in PIOsimHD

NameArtur Thiessen
AdvisorsDr. Julian Kunkel
Date2012-04-04

Report BibTeX Presentation

Auswirkungen verschiedener Parameter des LINPACK Benchmarks auf den Energieverbrauch

NameJohannes Pietrzyk
AdvisorsTimo Minartz
Date2012-02-19

Report BibTeX Presentation

Lokales Chat-System

NamePaul Lindt, Irina Lindt
AdvisorsDr. Michael Kuhn, Dr. Julian Kunkel
Date2011-12-15

Report BibTeX Presentation Sources

Lagrangesches Partikelausbreitungsmodell

NameJohann Weging, Cedrick Ansorge
AdvisorsDr. Julian Kunkel
Date2011-03-30
AbstractLaDis is a parallel implementation of a Lagrangian particle diffusion model. Particles sources like factories or streets emitting particles in to a pre computed wind field.
LaDis than calculate the path in which the particles will be distributed inside the wind filed. The area that is computed, is divided in a static grid which is scattered among the processes. Each process is holding multiple parts of the grid to add the ability of load balancing. The parallelization is done with MPI because of some performance problems with OpenMP. LaDis still has some bugs when it is passing the particles from one process to another.

Report BibTeX Presentation Project homepage

Implementierung einer Android-Applikation zur komfortablen Fernsteuerung eines Zielrechners

NameSenad Licina, Pascal Rost, Christopher Schewe
AdvisorsDr. Julian Kunkel
Date2011-09-28
AbstractDas Ziel des Projekts ist die Erstellung einer Android-Applikation zur Fernsteuerung eines Zielrechners. Dabei sollen sowohl Linux- als auch Windows-Systeme unterstützt werden. Es ist vorgesehen, dass sowohl der Maus-Cursor als auch Tastatureingaben mit dem Programm gesteuert werden können. Ebenfalls soll es möglich sein, vordefinierte oder vom Benutzer erstellte Befehle (Unix, Windows) oder auch Makros an den Zielrechner zu senden, wo diese ausgeführt werden. Für die auf dem Desktop häufig angewandten Handlungsschritte, wie beispielsweise dem Anzeigen aller aktiven Anwendungen oder dem Wechseln zwischen offenen Fenstern, sollen im Programm leicht auszulösende Aktionen existieren. Die Benutzeroberfläche der Anwendung soll sich dem jeweiligen Arbeitskontext anpassen lassen. Beispielhafte Anwendungskontexte sind der Umgang mit einem Multimediaprogramm wie VLC und das “Surfen” im Internet mittels eines Webbrowsers. Es sollen Profile für unterschiedliche Netzwerkkonfigurationen zur Verbindung des Mobiltelefons mit dem Zielsystem anlegbar sein. Programmkonfigurationen und Profile sollen exportierbar und zwischen unterschiedlichen Telefonen austauschbar sein. Wenn möglich, sollen alle Einstellungen des Programms sowie das Erstellen von Befehlen etc. auch mittels der Server-Applikation auf dem Zielsystem möglich sein (wenn eine Server-Applikation notwendig ist). Durch einen Synchronisationsprozess kommen die Daten auf das Telefon. Wenn möglich soll neben TCP/IP und SSH auch eine Verbindungsmöglichkeit mittels Bluetooth existieren.

Report BibTeX Project homepage

Moab Evaluation

NameFlorian Ehmke
AdvisorsTimo Minartz
Date2011-09-14
AbstractOften the energy costs are a large amount of the total costs of ownership (TCO) of a cluster. The longer the cluster is used the larger this slice of the cake grows. Different approaches exist to alleviate this problem. CPUs lower their clock speed when they are idle (which often already decreases the power consumption by 50 %), hard disks spin down and GPUs lower the clock speed of both core and memory. All of these methods are applied very fast and on demand. However the idle power consumption of today’s systems remains high. If large idle times are likely to happen the best option could be to completely shutdown a node. Most resource managers and schedulers don’t support this. As soon as one node is turned off it is reported as offline and no longer eligible to run jobs. Moab provides this functionality which is the topic of this report. We will investigate Moab’s behaviour during several different workloads and measure how much energy can be conserved.

Report BibTeX Presentation

Community Plattform for Parabench

NameDennis Runz, Christian Seyda
AdvisorsOlga Mordvinova, Dr. Julian Kunkel
Date2010-06-14
AbstractThe main goal of this practical was the development of a community website for Parabench, where users can share their IO patterns, their benchmark results including visualizations and a description of their testing plattform. To accomplish this, we choosed Django as our webframework to build upon, Highcharts as visualization tool and wrote a search interface based on machintags. In the following we will introduce the basics of Django, Highcharts, how we produce plots with both of them and how our search-backend is working.

Report BibTeX

Evaluating selected cluster file systems with programmable I/O-pattern benchmark Parabench

NameJens Schlager, Marcel Krause
AdvisorsDr. Julian Kunkel, Olga Mordvinova
Date2010-04-13
AbstractIn this report we’ll present in the part “Installation and experiences” a description of the chosen clustered file systems FhGFS, Tahoe and OCFS2 with their most interesting features. In this context we’ll explain the architecture of FhGFS and Tahoe. A second aspect there is the installation of these file systems and to write down the installation process for each of them. Occurring problems with these file systems in the installation process are also described. Because of problems with FhGFS and Tahoe we had to choose for the part “Evaluation with Parabench” two other runnable clustered file systems: Ceph and GlusterFS. We’ll also evaluate in this part OCFS2 that only remained from the part “Installation and experiences” with this new Benchmark. In our conclusion and summary we’ll describe the essential findings of the former two parts and we’ll give some forecasts about future research.

Presentation

Programmable I/O-Pattern Benchmark for Cluster File Systems

NameDennis Runz, Christian Seyda
AdvisorsOlga Mordvinova, Dr. Julian Kunkel
Date2009-11-25
AbstractThe main goal of this practical was to develop a benchmark for cluster file systems, where custom I/O-patterns can be passed to. To accomplish this, we designed an interpreter based benchmark. We use Flex and Bison to realize the scanner/parser layer for this approach. We also installed two modern cluster file systems, GlusterFS and Ceph, to keep up with current technologies. This document should explain the structure of the benchmark enough to understand the internals and to be able to work with the source code.

Report BibTeX

High Availability for the Remote File System

NameMichael Kuhn
AdvisorsDr. Julian Kunkel
Date2009-07-07

Report BibTeX Presentation Sources

Design & Implementation of the Remote File System

NameMichael Kuhn
AdvisorsDr. Julian Kunkel
Date2009-02-05

Report BibTeX Presentation Sources

PIOsim MPIWrapper to create XML-traces

NamePaul Müller
AdvisorsDr. Julian Kunkel
Date2009-04-30
AbstractThe MPI wrapper is a static library that can be used to log MPI function calls which occur during the execution of a program. The log files are intended to be used by the PIOsimHD simulator project. This document describes the use, internal structure and extension possibilities of the MPI wrapper.

Report BibTeX Presentation

Database Overlay File System with FUSE

NameTimo Minartz
AdvisorsDr. Julian Kunkel
Date2010-04-11
AbstractThe goal of this software project was to implement a lightweight filesystem with FUSE to map filesystem sources and database tables in one namespace for a multiuser environment.

Report BibTeX Presentation

Containerbasierte Metadatenverwaltung in PVFS2 für kleine Dateien

NameAndreas Beyer, Kai Husmann
AdvisorsDr. Julian Kunkel
Date2009-12-12
AbstractThis internship originated in the problem of optimising PFVS21 for small files through the integration of file meta data into directory entries. A PVFS independend solution for that problem, the so called container format, had been developed by Hendrik Heinrich already. Out of this the idea was born to implement a container feature (CF) into PVFS. Since it was not possible to integrade the H. Heinrich code into PVFS we decided to concentrate on its main issue: That is the sparing of redundant meta data. Unfortunately the speed evaluation of our code rather discourages from an implementation of a container feature for PVFS - at least if there is not planned a much higher degree afford for solving that problem. But with our design some states (within the client side state machines) can already be spared and with futher coding a reliable speed enhancement for containers will surely be possible. The meta data of the files within a container can be spared completely. It will be there only one time as meta data for the container itself. Therefore the smaller and numerous the files are, the bigger the enhancement considering hard disk usage. For a few big files this code will never be of much use - but it doesn't aim at that.

Report BibTeX

Evaluierung von Hadoop

NameTien Duc Dinh
AdvisorsDr. Julian Kunkel, Olga Mordvinova
Date2009-12-04
AbstractHadoop is an open-source, Java-based programming framework that supports the processing of large data sets in a distributed computing environment. It was inspired by Google MapReduce and Google File System (GFS) papers. Hadoop is now a top level Apache project, being built and used by a community of contributors from all over the world, since it’s easy to install, configure and can be run on many platforms supporting Java. The Hadoop framework is currently used by major players including Google, Yahoo and IBM,
largely for applications involving search engines and advertising. The major contributions of this work are a Hadoop performance evaluation on writing/reading and full understanding about the MapReduce concept as well as the distribution of processes. Section 3 describes the HDFS and MapReduce concept with a small example. Section 4 describes the installation and configuration a HDFS cluster. Section 5 describes the test preparation towards our cluster enviroment. Section 6 has performance measurements based on many test cases and a performance comparision between the HDFS and local file system on the testing cluster.

Report BibTeX Presentation

Analyse und Erweiterung der JavaGUI für PIOsimHD

NameDulip Withanage
AdvisorsDr. Julian Kunkel
Date2009-02-05
AbstractThis document describes the graphical user interface for the Parallel I/O Simulator(PIOsim) for clusters, developed by the Parallel and Distributed Systems Research Group of the Ruprecht-Karls Universität in Heidelberg, Germany. The graphical user interface is based on a previous implementation. The PIOsimGUI is a swing and awt based java application. The underlying modeling architecture Piosim which is also a java application was completely rewritten by Julian Kunkel during the internship period to facilitate the complex requirements of modeling the cluster. The integration of the model to the existing graphical user interface was executed by the author. New programming concepts were added to the programming model and to the graphical user interface in order to provide efficient and dynamical graphics rendering. The existing GUI was extended to facilitate more than one switch, component hosting machines that hold a set of components. A basic point-to-point connection drawing concept was introduced for connecting the cluster component. An easily modifiable attribute setting concept was added and the programming was executed in a one-to-one mapping of the model object for easily implementation of newer component types. Some of the restrictions of the earlier drawing area were eased as well. The document also discusses the technical difficulties that the author confronted in extending the graphical user interface. Most of the restrictions that were cased by the underlying swing and awt libraries are discussed and illustrated in examples. Furthermore, the document describes the concepts of the Model-View-Controller (MVC) architecture that is a generally accepted and mostly used concept of user interface designing. The relation of the current model to the MVC architecture is also discussed. The Eclipse Standard Widget Toolkit (SWT) is anaylsed in the last chapter. A sample SWT paint application is included as an executable jar file which comes with the eclipse SWT installation.

Report BibTeX Presentation

Evaluation von p2p Techniken für die E/A in verteilten Rechnerumgebungen

NameArne Babenhauserheide
AdvisorsDr. Julian Kunkel
Date2008-10-28
AbstractIn this report we analyzed potential usage of p2p networks for efficient data replication in distributed systems. Theoretical analysis of an optimal p2p replication structure shows a huge potential for improvements in the speed of data replication using a distributed system where each node in the network uploads and downloads with full speed at the same time. We gathered and evaluated data in 7 nodes of our test cluster interconnected by a 1Gbps network and for a simulated WLAN environment. However measured performance sticks behind theoretical considerations.

Report BibTeX Presentation

Evaluation of FUSE with a Memory File System

NameMichael Kuhn
AdvisorsDr. Julian Kunkel
Date2008-10-12

Report BibTeX Presentation Sources

Container Library and FUSE Container File System

NameMichael Kuhn
AdvisorsDr. Julian Kunkel
Date2008-03-31

Report BibTeX Presentation Sources

Neighbourcast, a Network Paradigm for MMORPGs

NameAlexei Bratuhin
AdvisorsDr. Julian Kunkel, Prof. Dr. Thomas Ludwig
Date2008-09-30
AbstractIn MMORPGs the most common information exchanged is the status, the position and the stats information. This information needs to be spread all over the virtual world of MMORPG. But no one from one corner/pole of the map needs to know what’s happening on the other corner/pole (unless this is a system or some kind of multicast (group/guild) event). Therefore information about world changes needs to be spread among the ’neighbours’ of the change only. The most reasonable metric for neighbourhood relation in MMORPG is the sight range of a player. Develop an algorithm/model for MMORPG that uses position/range information for p2p communication, thus avoiding broadcast for information exchange between nodes that leads to network overloading.

Report BibTeX Presentation

Benchmarking von XEN-virtualisierten Servern

NameArne Klein
AdvisorsDr. Julian Kunkel
Date2008-03-01

Report BibTeX Presentation

Java-GUI für die Modellierung eines Clusters im PIOsim

NameMathias Braun
AdvisorsDr. Julian Kunkel, Prof. Dr. Thomas Ludwig
Date2007-10-14
AbstractDiese Ausarbeitung behandelt die Entwicklung einer grafischen Oberfläche (GUI) eines Simulators für parallele Ein-/Ausgabe, welcher zeitgleich von zwei anderen Studenten entwickelt wurde. Es wird hierbei hauptsächlich auf die Implementierung in Java und die Struktur der Benutzeroberfläche eingegangen. Ebenso werden die Möglichkeiten zur Erweiterung und Weiterentwicklung der GUI diskutiert. Die Arbeit wird durch Beispiele zur Bedienung der Oberfläche sowie anschauliche Abbildungen abgerundet.

Report BibTeX

Simulations-Prototyp für PIOsim

NameAnton Ruff, Artjom König
AdvisorsDr. Julian Kunkel, Prof. Dr. Thomas Ludwig
Date2007-10-09

Verwendung von TAU Trace und OTF zur Leistungsanalyse

NameSteffen Janz
AdvisorsDr. Julian Kunkel
Date2007-10-08

Presentation

Evaluierung paralleler Dateisysteme

NameMichael Kuhn, Christian Lohse
AdvisorsProf. Dr. Thomas Ludwig
Date2006-09-12

Report BibTeX Presentation