86 research outputs found
North, South, East and West: Best Practices in Information Literacy Services for International Students at Arizona Universities
No full textabstract: Poster about meeting the academic and cultural needs of international students at the Arizona State University Libraries and the University of Arizona Libraries. The poster presentation focuses on: (1) strategies to promote information literacy skills of international students in the two university libraries, (2) what the libraries are doing to improve services to meet the needs and encourage library use among international students; and (3) partnerships that have been established with other academic departments or institutions.Poster presented at the 2014 Arizona Library Association annual conference (2014 AzLA/MPLA Conference)
Zeitreihenanalyse auf dünnen Gittern
No full textZeitreihen sind Mengen von zeitlich geordneten Beobachtungen und fallen bei nahezu allen messbaren Daten an. In dieser Arbeit wird das Vorhersageproblem für Zeitreihen untersucht, für das viele praktische Anwendungen existieren, darunter die Vorhersage von Börsendaten. Für die Untersuchung von Zeitreihen können Gitter-basierte Ansätze verwendet werden. Bei diesen treten jedoch bei hohen Problemdimensionen unpraktikabel große Rechenzeiten auf. In dieser Arbeit wird eine Methode zur Zeitreihenanalyse mit dünnen Gittern vorgestellt, die es erlaubt, Lösungen für Probleme mit höherer Dimensionalität zu berechnen. Die durchgeführten Experimente zeigen dabei, dass für einige Datensätze Vorhersagen mit sehr hoher Qualität berechnet werden. Gleichzeitig ist die benötigte Rechenzeit für viele zeitkritische Anwendungen bereits ausreichend. Um das Anwendungsspektrum der Methode weiter zu vergrößern, werden Optimierungen vorgestellt, mit denen die benötigte Rechenzeit weiter verringert wird
Auto-tuning and performance portability on heterogeneous hardware
No full textIn high-performance computing, excellent node-level performance is required for the efficient use of supercomputers. However, manual optimization is a tedious process that commonly needs to be repeated for every hardware platform targeted. Auto-tuning has been developed as an approach to partially automate the optimization process, generally by tuning parameterized compute kernels. Through auto-tuning, both high performance on a single hardware platform and performance portability can be achieved.
In this work, we present the auto-tuning framework AutoTuneTMP that leverages two features: just-in-time (JIT) compilation and C ++ template metaprogramming. JIT compilation enables a tight integration of auto-tuning into an application and reduces the time required for auto-tuning. Template metaprogramming enables a design focused on ease-of-integration, maintainability and extensibility. It further forms the basis for optimization templates that support the development of auto-tunable compute kernels. Additionally, our framework provides search strategies for performing parameter tuning.
To demonstrate the applicability and usefulness of our framework, we use Auto-TuneTMP to auto-tune three algorithms that require excellent performance. The first algorithm is matrix multiplication. Whereas many implementations are manually optimized or even use assembly, we rely on template metaprogramming for a high-level approach. Across four hardware platforms, we achieved up to 91% of the peak performance and are, therefore, competitive with vendor libraries such as Intel’s MKL.
Sparse grid regression is well-suited for machine learning in moderate-dimensional big data scenarios. However, large datasets necessitate high-performance algorithms. We introduce two auto-tuned high-performance algorithms for this application: the unified streaming algorithm and the subspace algorithm. These algorithms serve as the second and third example for auto-tuning with AutoTuneTMP. The unified streaming algorithm is written in OpenCL and targets a wide range of architectures including GPUs. The subspace algorithm was developed for processor platforms only, but has a lower time complexity. Due to these auto-tuned algorithms and a new approach for the spatial adaptivity of sparse grids, speedups of up to 13x were measured compared to the state-of-the-art surplus-refined masked streaming approach. Furthermore, we demonstrate that both new algorithms are performance-portable.
Apart from auto-tuning, we investigate the performance portability of a new distributed variant of the sparse grid clustering method. As a density-based clustering method, it relies on a sparse grid density estimation to compute density functions for large datasets of moderate dimensionality in linear complexity in the size of the dataset. The algorithm consists of four major components: two density compute kernels and two compute kernels for creating and pruning a k-nearest-neighbor graph. All components were written using OpenCL and MPI. On the node-level, we reached on average 79% of the achievable peak performance of one processor and four GPUs. In distributed experiments on two supercomputers, Hazel Hen and Piz Daint, we measured up to 352 TFLOPS using 128 nodes. As a similar fraction of peak performance was achieved on both supercomputers and because of the high node-level efficiency, we can demonstrate the performance portability of the algorithm.
Our work shows that performance portability is a realistic goal for scientific applications on modern hardware. By using JIT compilation and template metaprogramming, the tightly-integrated auto-tuning approach presented reduces the effort required for optimization without compromising on performance
Eine künstliche Intelligenz für das Kartenspiel Tichu
No full textTichu ist ein Kartenspiel, bei dem klassische Suchalgorithmen aufgrund der Unkenntnis der Verteilung der Karten nicht effizient angewendet werden können. In dieser Arbeit wird eine künstliche Intelligenz vorgestellt, die trotz dieser Schwierigkeit stark spielt und alle Entscheidungen in kurzer Zeit trifft. Dies wird mit einem einfachen Suchalgorithmus und einer komplexen Bewertungsfunktion erreicht, wobei der Bewertungsfunktion insbesondere ein Modell der Initiative einer Aktion zugrunde liegt. Die mit diesem Ansatz erzielte Spielstärke der künstlichen Intelligenz ist mit der Spielstärke durchschnittlicher menschlicher Spieler vergleichbar. Es zeigt sich, dass mittels komplexer, hinter der Bewertungsfunktion stehender Modelle auf einen Suchalgorithmus mit hoher Suchtiefe bei Tichu verzichtet werden kann
Sparse Finite Gabor Frames for Operator Sampling
No full textPublication in the conference proceedings of SampTA, Bremen, Germany, 201
Robust Replication Control Is Generated by Temporal Gaps between Licensing and Firing Phases and Depends on Degradation of Firing Factor Sld2
No full textSummaryTemporal separation of DNA replication initiation into licensing and firing phases ensures the precise duplication of the genome during each cell cycle. Cyclin-dependent kinase (CDK) is known to generate this separation by activating firing factors and at the same time inhibiting licensing factors but may not be sufficient to ensure robust separation at transitions between both phases. Here, we show that a temporal gap separates the inactivation of firing factors from the re-activation of licensing factors during mitosis in budding yeast. We find that gap size critically depends on phosphorylation-dependent degradation of the firing factor Sld2 mediated by CDK, DDK, Mck1, and Cdc5 kinases and the ubiquitin-ligases Dma1/2. Stable mutants of Sld2 minimize the gap and cause increased genome instability in an origin-dependent manner when combined with deregulation of other replication regulators or checkpoint mechanisms. Robust separation of licensing and firing phases therefore appears indispensable to safeguard genome stability
Author response
No full textHolliday junctions (HJs) are key DNA intermediates in homologous recombination. They link homologous DNA strands and have to be faithfully removed for proper DNA segregation and genome integrity. Here, we present the crystal structure of human HJ resolvase GEN1 complexed with DNA at 3.0 Å resolution. The GEN1 core is similar to other Rad2/XPG nucleases. However, unlike other members of the superfamily, GEN1 contains a chromodomain as an additional DNA interaction site. Chromodomains are known for their chromatin-targeting function in chromatin remodelers and histone(de)acetylases but they have not previously been found in nucleases. The GEN1 chromodomain directly contacts DNA and its truncation severely hampers GEN1's catalytic activity. Structure-guided mutations in vitroand in vivo in yeast validated our mechanistic findings. Our study provides the missing structure in the Rad2/XPG family and insights how a well-conserved nuclease core acquires versatility in recognizing diverse substrates for DNA repair and maintenance
Computational Efficiency vs. Maintainability and Portability. Experiences with the Sparse Grid Code SG++
No full text- …
