In this bimonthly feature, HPCwire highlights newly published research in the high-performance computing community and related domains. From parallel programming to exascale to quantum computing, the details are here.
Charting the race to exascale in Europe
The European Union has been doubling down on its commitment to HPC with initiatives like EuroHPC. In this paper, authors from the Polytechnic University of Catalonia and the Sapienza University of Rome chart this European race toward exascale, covering the politics, the economics, recent events in HPC architecture development, and the European policy initiatives that have been announced.
Authors: Fabrizio Gagliardi, Miquel Moreto, Mauro Olivieri and Mateo Valero.
Using multi-level parallelism on HPC applications hosted on Azure
Cloud computing is an extremely cost efficient way to engage with HPC applications. In this paper, by a team from the Informatics Research Institute in Egypt, the authors evaluate the impact of using multi-level parallelism on intensive parallel tasks hosted on a virtualized HPC cluster. With an eye toward performance cost, they run a set of experiments, concluding that balancing workload between processes and threads per process are key to efficiency.
Authors: Hanan A. Hassan, Mona S. Kashkoush, Mohamed Azab and Walaa M. Sheta.
Evaluating the Arm ecosystem for HPC
Arm-based processors have been shaking up the HPC scene, offering an alternative to x86 processors. In this paper, a team from the University of Edinburgh evaluates the hardware and software offerings of the Arm ecosystem for HPC use, benchmarking a production HPC platform based on ThunderX2 processors. The authors conclude that performance is as good as – or better than – well-established platforms.
Authors: Adrian Jackson, Andrew Turner, Michele Weiland, Nick Johnson, Olly Perks and Mark Parsons.
Reproducing scientific workflows at extreme scales
As highlighted in the previous edition of What’s New in HPC Research, reproducing complex models at extreme scales can pose serious scaling difficulties. In this paper, written by a team from a number of national laboratories, the authors propose an approach for improved reproducibility. Focusing on capturing and relating provenance characteristics and performance metrics, they evaluate their approach on earth system models and molecular dynamics workflows.
Authors: Line Pouchard, Sterling Baldwin, Todd Elsethagen, Shantenu Jha, Bibi Raju, Eric Stephan, Li Tang and Kerstin Kleese Van Dam.
Exploring the design space of next-generation HPC machines
As the landscape of HPC architectures expands due to new technology and increasing complexity, tools are required for analyzing and predicting the impact of new architectural features. In this paper, written by a team from the Barcelona Supercomputing Center, the authors simulate five hybrid applications over 864 state-of-the-art architectural proposals. Using the results from these simulations, they provide architectural recommendations to serve as hardware and software design guidelines.
Authors: Constantino Gomez, Francesc Martinez, Adria Armejach, Miquel Moreto, Filippo Mantovani and Marc Casas.
Automating user registration on academic HPC systems
The demand for HPC cluster systems is growing, but the time necessary to register users for academic clusters can cause long delays in usability. In this paper, authors Junya Nakamura and Masatoshi Tsuchiya propose an academic HPC cluster system with automated user registration. This new method, the authors suggest, will allow users to to utilize academic clusters on-demand rather than encountering delays.
Authors: Junya Nakamura and Masatoshi Tsuchiya.
Studying the use of TensorFlow for machine learning on state-of-the-art HPC clusters
As the development of specific architectures rapidly develops, there is growing uncertainty as to which are the best hardware/software configurations to efficiently support specific tasks like machine learning. In this paper, a team from the Barcelona Supercomputing Center examines the workflow of TensorFlow for image recognition. The authors highlight the dependency of the performance in the training phase on the availability of specific libraries and test TensorFlow on several architectures.
Authors: Guillem Ramirez-Gargallo, Marta Garcia-Gasulla and Filippo Mantovani.
Do you know about research that should be included in next month’s list? If so, send us an email at [email protected]. We look forward to hearing from you.