Historie 2012
6.Dezember 2012: Christian Pflüger (University of Utah, Huntsman Cancer Institute) "Hacking the Epigenome: Reprogramming at its best." (Folien)
Human stem cells are thought to be one of the most promising tools in medicine. They can be used to treat people with detrimental diseases that require new healthy tissue in order to stop the symptoms or even to cure it (e.g. parkinson disease). Recent advances in stem cell biology created opportunities to investigate stem cell formation from adult differentiated tissue such as skin (fibroblasts) or blood. These new methods enable biologist to shorten the time period for induced stem cell formation from months to mere days with high efficiency. This in turn enables high throughput analysis using next-generation-sequencing (NGS) of transcripts (e.g mRNA) and various epigenetic modifications such as DNA methylations and DNA demethylation. The resulting massive amounts of information and data require sophisticated bioinformatic analysis to understand key regulatory elements and regions during iPS formation. This talk aims at pointing out the current state of the art bioinformatic analysis methods and offers a wish list for informatics people from a biologist's perspective.
22. November 2012: Ruedi Seiler (TU Berlin) "Das Q-Sanov Theorem, eine Krippe fundamentaler Resultate der Informationstheorie"
Das Quanten-Sanov Theorem und die darin verwendeten Begriffe werden vorgestellt und daraus drei fundamentale asymptotische Resultate der Informationstheorie hergeleitet und diskutiert:
- Der Satz von Shannon
- Der Satz von Kaltchenko und Yang.
- Das Steinschen Lemma
Grundlegend für alle Resultate ist der Begriff typischer Mengen, wie wir das aus dem Zufallsexperiment Münzwurf kennen: Die Folge der Ergebnisse enthält ungefähr soviele Köpfe wie Zahlen.
25. Oktober 2012: Torsten Höfler (ETH Zürich) "New Features in MPI-3.0 in the Context of Exascale Computing" (Folien)
The new MPI-3.0 standard sets out to address the changes in technology that happened in the last years. This modernization of MPI also addresses architectures for future large-scale computing systems and changed user-needs. The MPI Forum introduced new collective communication interfaces, such as nonblocking and neighborhood collectives, better support for topology mapping, multicore computers, and a largely extended one sided access model. This talk will provide an overview of those new features in the context of the changing environments and addresses strengths and remaining weaknesses of MPI. It can be expected that all those features are quickly available in many MPI implementations and thus influence the state of the art in parallel programming.
2. August 2012: Robert Henschel (Indiana University): "Die IU/ZIH Kollaboration - ein Überblick" (Folien)
Die Kollaboration zwischen Research Technologies (RT), einer Abteilung der Indiana University, und dem Zentrum für Informationsdienste und Hochleistungsrechnen (ZIH), einer zentralen wissenschaftlichen Einrichtung der TU Dresden, wurde 2006 durch Craig Stewart initiiert und 2008 durch dass Unterzeichnen eine Absichtserklärung (MoU) formalisiert. Das MoU beschreibt vier Schwerpunkte für die Zusammenarbeit der beiden Institute: Datenintensives Rechnen, Hochleistungsrechnen in der Bioinformatik, Dateisysteme in Weiterverkehrsnetzen und Performanceanalyse von wissenschaftlichen Anwendungen. Der Vortrag behandelt die gemeinsamen Arbeiten innerhalb des FutureGrid-Projektes sowie die Partnerschaft im 100GBit-Projekt der TU Dresden und gibt einen Einblick in die interkontinentale Zusammenarbeit.
26. Juli 2012: Lucas Schnorr (CNRS Grenoble/Frankreich): "Data Aggregation and Alternative Visualization Techniques for Parallel and Distributed Program Analysis" (Folien)
One of the main visualization techniques used for the analysis of parallel program behavior is the space-time view, or Gantt-chart. The visualization scalability of this technique is usually limited by the screen size and the number of processes that can be represented on it. This talk presents a combination of data aggregation techniques and different ways of visualizing trace information, achieving a better visualization scalability on the analysis. It details the spatial and temporal data aggregation performed on the traces, and then presents two interactive aggregation-enabled visualization techniques: the squarified treemap and the hierarchical graph. Application scenarios are used to illustrate these techniques in practice.
19. Juli 2012: Thomas Lippert (FZ Jülich): Das europäische Exascale-Projekt DEEP - Auf dem Weg zur "Dynamical Exascale Entry Platform"
With begin of 2012, a consortium of 16 partners from 8 countries led by the Jülich Supercomputing Centre, among them 5 industrial partners, is engaged in developing the novel hybrid supercomputing system DEEP. The DEEP project is funded by the European Community under FP7-ICT-2011-7 as Integrated Project, with co-funding by the partners. The DEEP concept foresees a standard cluster computer component complemented by a cluster of accelerator cards, called booster. DEEP is an experiment with the aim to adapt the hardware architecture to the hierarchy of different concurrency levels of application codes. Due to the cluster-booster concept, for a given code, cluster as well as booster resources can be assigned to different parts of the code in a dynamical manner, thus optimizing scalability. This is achieved through an adaption of the cluster operating software ParaStation (ParTec) along with the parallel programming environment OmpSS (BSC). The major challenge for the concept is to achieve a proper and most efficient interaction between cluster and booster while minimizing the communication between both parts. Moreover, it is the combination of Intel's Many Core Integrated Architecture (MIC, Intel Braunschweig) and the EXTOLL communication system (Uni Heidelberg) that allows to boot the booster cards without additional processor. DEEP promises an unprecedented performance, scalability as well as energy efficiency of the booster system. Energy efficiency is further improved through hot water cooling technology (LRZ, EuroTech). Six European partners contribute with porting their applications that all exhibit several concurrency levels and are expected to require Exascale performance in the future.
28. Juni 2012: Bertil Schmidt (Uni Mainz) "Parallel Algorithms and Tools for Bioinformatics on GPUs" (Folien)
High-throughput techniques for DNA sequencing have led to a rapid growth in the amount of digital biological data. The current state-of-the-art technology produces 600 billion nucleotides per machine run. Furthermore, the speed and yield of NGS (Next-generation sequencing) instruments continue to increase at a rate beyond Moore's Law, with updates in 2012 enabling 1 trillion nucleotides per run. Correspondingly, sequencing costs (per sequenced nucleotide) continue to fall rapidly, from several billion dollars for the first human genome in 2000 to a forecast US$1000 per genome by the end of 2012. However, to be effective, the usage of NGS for medical treatment will require algorithms and tools for sequence analysis that can scale to billions of short reads. In this talk I will demonstrate how parallel computing platforms based on CUDA-enabled GPUs, multi-core CPUs, and heterogeneous CPU/GPU clusters can be used as efficient computational platforms to design and implement scalable tools for sequence analysis. I will present solutions for classical sequence alignment problems (such as pairwise sequence alignment, BLAST, multiple sequence analysis, motif finding) as well as for NGS algorithms (such as short-read error correction, short-read mapping, short-read clustering).
24. Mai 2012: Marc Casas Guix (LLNL) "Automatic Phase Detection and Structure Extraction of Parallel Applications" (Folien)
Tracing is an accepted and well-known approach to understand and improve the performance of high performance computing applications. However, generating and analyzing trace-files obtained from large scale executions can be really problematic due to the large amount of data generated by such massively parallel executions. Thus, automatic methodologies should be applied to reduce the size of the data, ruling out its non-significant or redundant parts and keeping the fundamental ones. In this talk, a solution based on signal processing techniques, Wavelet and Fourier transforms, will be presented. By analyzing the specter of frequencies that appear in applications’ executions, the approach is able to detect the internal structure of parallel executions and to rule out redundant information, reducing by one or two orders of magnitude the data that should be analyzed. Finally, more general considerations regarding high performance computing and the challenges that exascale computing brings will also we made.
26.April 2012: Thomas Cowan (Direktor des Instituts für Strahlenphysik am Helmholtz-Zentrum Dresden-Rossendorf (HZDR)) "Beschleunigung der Beschleunigung - Lasergetriebene Strahlungsquellen und ihre Anwendungen"
Neuartige Strahlungsquellen ermöglichen nicht nur neue Einblicke in ultraschnelle Vorgänge in Materie sondern auch deren Kontrolle. Eine derartige Kontrolle erfordert genaues Wissen über die Erzeugung von elektromagnetischer wie Teilchen-Strahlung sowie deren Wechselwirkung mit Materie auf atomarer Ebene. Um dieses Wissen zu vergrößern braucht es eine enge Verbindung von experimentellen Messungen, Datenanalyse und Simulation.
Der Vortrag stellt neuartige Strahlungsquellen vor und diskutiert ihre Bedeutung in der fundamentalen Forschung ebenso wie ihre zukünftige Anwendung, zum Beispiel in der Krebstherapie. Als Beispiele aktueller Forschung dienen die Beschleunigung von Teilchenstrahlen mit Hilfe von Lasern und die in-vivo Dosimetrie bei der Krebsbehandlung mit Ionenstrahlen. Beide Forschungsgebiete profitieren von der Beschleunigung komplexer Rechenoperationen durch GPUs, sowohl im Bereich der Simulation als auch in der Datenauswertung.
22. März 2012: Josef Weidendorfer (TU München): "Architecture Simulation for Programmers"
To study performance bottlenecks of (parallel) programs, analysis tools usually take advantage of a mix of hardware performance counters and application instrumentation as event source. Real hardware properties are measured, showing details about the symptoms of any performance problem. However, this real view to hardware can be tricky: for the tool, as instrumentation overhead can invalidate the measurement; and for the user, as event types can be difficult to interpret. Architecture simulation can overcome these obstacles and provide more abstract metrics not measurable in legacy processor hardware.
This talk will focus on using cache simulation for detailed analysis of memory access behavior of programs, and show the benefits of this approach, such as better abstract metrics than just hit/miss ratios for cache exploitation. In this regard, upcoming extensions to the tool suite Callgrind/KCachegrind are shown, as well as research on keeping the simulation slowdown small.
23. Februar 2012: Martin Hofmann-Apitius (Fraunhofer SCAI) "Large-Scale Information Extraction for Biomedical Modelling and Simulation" (Folien)
Unstructured information is a huge resource for scientific information. This is in particular true for sciences with a strong empirical background, such as biology, pharmaceutical chemistry or medicine. In my talk, I will give an overview on our work that aims at making scientific information available that is "hidden" in scientific publications (including patents) and medical narratives (electronic patient records). The presentation will cover essentials of information extraction technologies developed in our lab, their implementation in workflows for large-scale production of relevant information and the application of our information extraction technologies in the area of modelling neurodegenerative diseases.
26. Januar 2012: Michael Hohmuth (AMD, OSRC ): "OS and Architecture Research at the AMD Operating System Research Center: ASF, the Advanced Synchronization Facility"
In this talk, I will present the Advanced Synchronization Facility (ASF), an experimental AMD64 architecture extension aimed at one of these trends, parallel computing. ASF is designed to make parallel programming easier by supporting two styles of writing parallel programs: lock-free programming and transactional programming.