January 2, 2013

Polish HPC Consortium Boosts Prospects for Local Scientists

Nicole Hemsoth

Poland is not usually thought of as a supercomputing powerhouse. Until recently, most of the local research and academic centers housed only modest-sized HPC clusters for Polish researchers. That is now changing with the POWIEW project, a consortium devoted to bringing world-class high performance computing to the nation’s scientists.

POWIEW provides five state-of-the art systems, based on the latest processor technology from Intel, AMD, IBM, and NVIDIA. The machines are spread across the consortium’s three member organizations: the Interdisciplinary Centre for Mathematical and Computational Modeling (ICM) in Warsaw, Krakow’s Academic Computer Centre (CYFRONET), and the Poznan Supercomputing and Networking Center (PSNC).

We asked Maciej Filocha, POWIEW’s project manager and director of its HPC division, to describe the impetus behind the organization and how they serve local researchers and scientists.

HPCwire: Can you describe the mission of POWIEW and the rationale for developing an indigenous supercomputing project for Poland?

Maciej Filocha: The POWIEW project sets up a unified strategy concept for the HPC development in Poland addressing research and academia. It encompasses a program of significant computational infrastructure enhancement, associated R&D programs and graduate and postgraduate education.

Since the early 2000s, the Polish HPC ecosystem was dominated by capacity-level cluster solutions — still relevant for many scientists, but obviously less so for “real” large-scale parallelism. To ensure competitiveness and attractiveness of Polish research organizations, local computational infrastructure was to be enhanced accordingly.

HPCwire: How do Polish scientists and researchers view the significance of HPC for their work?

Filocha: The POWIEW consortium consists of three leading research and academic HPC centers, founded about 20 years ago. From the beginning, the number of users has been growing and includes representatives of “classical” computational and life sciences as well as material sciences, engineering and environmental sciences.

HPCwire: What’s the level of funding for the project and where does it come from?

Filocha: The project is founded by the EU from the Innovative Economy Programme (85 percent), complemented by domestic sources (15 percent), with total budget of about 23 million euros (17 million USD). This covers not only acquisition and deployment of new computing systems but also R&D activities related to porting and optimization of selected codes for new architectures and general enabling actions.

HPCwire: What supercomputers are currently up and running at POWIEW, and how are they being used?

Filocha: The POWIEW Project focuses on two major HPC application areas: massively parallel processing (MPP) that delivers high scalability in fine-grained parallelism and symmetric multiprocessing for intensively coarse-grain parallel computational applications.

For the MPP class, the Blue Gene/P solution was chosen and has been running for almost two years with a full system load. The Blue Gene/P system is maintained by the CM) in Warsaw. This architecture is particularly useful by material science and life science researchers for its performance with high scalability on MPI applications. The system is also heavily used by neuroscientists where it enables simulations of large neural networks with reasonable performance. Most of the jobs running on the Blue Gene/P utilize a few thousand CPU cores.

For the second class of systems two solutions were identified: the IBM Power775 system installed at ICM in Warsaw and the SGI Altix UV SMP machine at the PSNC in Poznan. Both systems use the fat-node approach with the SGI UV implementing “true” SMP — one super node — while the Power775 represents a cluster of super nodes.

The Power775 machine is currently in operation for almost one year. It is used for the most demanding workloads, including high resolution atmosphere studies for weather predictions and very large cosmological simulations. The system has proven its high performance for memory-intensive and computing-intensive tasks.

The SGI machine in Poznan, being also a PRACE Tier-1 site, is the only SMP system of comparable size in Poland. It is used for memory-intensive tasks including reservoir modeling and complex simulations in astrophysics.

As a third technology choice, hardware-accelerated clusters have been identified. Both HPC centers in Poznan, and CYFRONET in Cracow, installed accelerated clusters choosing GPGPU solutions. GPUs in Cracow constitute a part of the largest supercomputer in Poland. Accelerators are widely utilized there to optimize locally developed codes for quantum chemistry computations and complex dynamics in astrophysics applications.

Another installation is the SGI/Rackable system maintained by PSNC, using AMD x86 servers accelerated with NVIDIA GPU cards. It is used for computing intensive tasks in molecular modeling and fluid flow dynamics in porous media — reservoir modeling.

HPCwire: Why such a wide variety of architectures? Doesn’t that create problems for users who want to share applications across platforms?

Filocha: An underlying idea for POWIEW was to provide all existing key HPC architectures based on complementarity and competencies sharing among project partners. Project experts are expected to provide support to the researchers so as optimize their choice of suitable architectures.
HPCwire: Poland is a member of PRACE. How does POWIEW fit into that consortium?

Filocha: All POWIEW project members are actively involved in PRACE activities since their beginning. They work in applications, hardware and policy-related tasks. Selected systems deployed within POWIEW project are now included as a Tier-1 systems in current Distributed European Computing Initiative (DECI) calls. Some of our computers, including the IBM 775 system are the first of its kind available for PRACE users. Our experience of day-to-day use of such systems allowed us to contribute significantly to best practices guides for PRACE users.

HPCwire: As far as the future of POWIEW, what’s being planned: new systems, collaborations, new application areas…?

Filocha: Formally, POWIEW will run until mid-2013, but the actual goal is to extend the deployed hardware infrastructure and acquired software competencies further, based on experience gained during last three years of intensive growth.