Scalable Infrastructure for the Automated Performance Analysis of Parallel Codes (SILC)

Overview

Performance analysis tools are an important part of the HPC ecosystem, because non-uniform and heterogeneous hardware platforms with complex software components cannot be optimized for computational efficiency, performance, or scalability without detailed insight into the parallel run-time behavior and guidance towards computational hot-spots and performance bottlenecks.

There are a number of established tools for this purpose. All use the same fundamental data recording techniques, namely profiling generation and event trace recording. On top of this, they perform different and complementary analysis methods.

The main goal of the SILC project was to design and implement a joint instrumentation and run-time data collection infrastructure for the established performance analysis tools Periscope, Scalasca, TAU, and Vampir. This provides a three-fold advantage. First, it enhances the interoperability between the tools. Second, it removes redundancies for maintenance, support, user training, and the implementation of new features in the future. This includes the support for new parallel programming methods or novel HPC platforms. It will also allow to share the effort between the participating groups and to use the development resources more efficiently. And third, it will relief the users of said performance analysis tools from redundant steps during experiment preparation and run-time data collection. In the past, every tool used different commands and options for the essentially same steps. Now, the joint infrastructure provides one way to produce measurement data that can be used for all four tools.

Near the end of the project funding period, the "Score-P" instrumentation and run-time measurement system is available as Open Source software and works together with the project partner's performance analysis tools. Furthermore, all partners are committed to a long-term collaboration beyond the funding period for this essential common software components and are open for new partersto join.

Partners & Associated Partners

Partners:

• Jülich Supercomputing Centre at Forschungszentrum Jülich
• the Center for Computing an Communication at RWTH Aachen
• Chair for Architecture of Parallel and Distributed Systems at Technische Universität München
• the GNS (Gesellschaft für numerische Simulation) mbH Braunschweig
• the Center for Information Services and HPC at Technische Univeristät Dresden (project coordinator)

Associated partners:

• Laboratory for Parallel Programming at the German Research School for Simulation Sciences, GRS-SIM
• GWT-TUD (Gesellschaft für Wissens- und Technologie-Transfer) Dresden
• Performance research Lab at the University of Oregon

The Periscope tool out is developed at TU München, Scalasca is developed at the Jülich Supercomputing Centre and at GRS-SIM, TAU is developed at the University of Oregon, and Vampir is developed at TU Dresden.

Furthermore, the SILC project is carried out in close cooperation with the PRIMA project, a collaborative project between the University of Oregon and Forschungszentrum Jülich funded by the US Department of Energy.

Project Results

As the most important results of the SILC project, the following software components are released

• Score-P (Scalable Performance Measurement Infrastructure for Parallel Codes) which provides the central functionalities for code instrumentation and run-time data collection.
• The Open Trace Format Version 2 (OTF2) is the highly-scalable event-trace data format together with a support library.
• CUBE4 is a highly-scalable profiling data format together with a support library and the CUBE4 GUI.
• The Opari2 source-to-source instrumentor for OpenMP codes.

All components are made available under a New-BSD Open Source license, see the resources section. For a more detailed overview about components, features and functionalities see the included documentation and the referenced presentations.

Also, the performance analysis tools Periscope, Scalasca, TAU, and Vampir have been adapted to the new infrastructure. Furthermore, a large number of tests with benchmarks and application test cases from academia and industry have been conducted.

Impact on HPC Community

The new common tools platform will provide a stable infrastructure for a number of performance analysis tools. This will increase the interoperability between the tools and improve the user experience when using several tools in combination. Since maintenance, support, and functionality extensions for the Score-P infrastructure will be carried in a collaborative manner even beyond the project period, there will be more combined development resources and less redundant effort in the future. This will allow faster adaptation of new important HPC trends and lower effort per group for development, support, testing, and training by distributing this tasks between all groups. In the future, the Score-P community effort is also open for new partners.

Already, a number of follow-up projects which extend and re-use the Score-P infrastructure created during the SILC BMBF project.

• LMAC (Performance dynamics of massively parallel codes), BMBF HPC Call 2, 08/2011-07/2014 (http://gauss-allianz.de/en/projects/2call.html → LMAC)
• H4H (Hybrid Parallel Programming to Optimise HPC Applications on Heterogeneous Architectures), ITEA2 project, 2010-2013,
• Hopsa (HOlistic Performance System Analysis), EU-Russia project, 02/2011-01/2013,
• GASPI (Global Address Space Programming Interface) , BMBF HPC Call 2, 06/2011-05/2014, (http://gauss-allianz.de/en/projects/2call.html → GASPI)

Resources

Software packages, documents, presentation slides, and URLs

• The project webpage is available at http://www.score-p.org, here, always the most recent release versions will be available
• A beta release version from SC'11 is available for download at http://www.vi-hps.org/datapool/page/41/scorep-1.0-beta.tar.gz or locally, a release version will follow in December 2011.
• Slides and paper preprint from the CI-HPC Conference in June 2010 in Schwetzingen: SILC_CiHPC_2010-06_slides.pdf and SILC_CiHPC_2010-06_paper_preprint.pdf
• Slides from the OTF2 BOF in Nov. 2010 at SC'10 in New Orleans (slides)
• Slides from the BMBF Status Meeting Nov. 2010 in Berlin (slides)
• Slides and paper from 2011 Dresden Tools Workshop (paper, slides)
• Slides from the Score-P BOF in Nov. 2011 at SC'11 in Seattle (slides)

References and contact information

The SILC project is funded by the German Bundesministerium für Bildung und Forschung (BMBF) from 01.01.2009 until 31.12.2011.

Dr. Andreas Knüpfer
Dresden University of Technology
Center for Information Services and High Performance Computing (ZIH)
D-01062 Dresden

Phone +49 351 463 38323
Emaill: This email address is being protected from spambots. You need JavaScript enabled to view it.