January 5 – 10 , 2014, Dagstuhl Perspectives Workshop 14022
Connecting Performance Analysis and Visualization to Advance Extreme Scale Computing
1 / 2 >
For support, please contact
Dagstuhl Report, Volume 4, Issue 1
Dagstuhl Manifesto, Volume 5, Issue 1
Aims & Scope
List of Participants
Dagstuhl's Impact: Documents available
Dagstuhl Perspectives Workshop Schedule [pdf]
Over the last decades an incredible amount of resources has been devoted to building ever more powerful supercomputers. However, exploiting the full capabilities of these machines is becoming exponentially more difficult with each new generation of hardware. In the systems coming online at this moment, application developers must deal with millions of cores, complex memory hierarchies, heterogeneous system architectures, high-dimensional network topologies as well as a host of other hardware details that may effect the performance of a code. To help understand and optimize the behavior of massively parallel simulations a new subfield of computer science has grown devoted to developing tools and techniques to collect and analyze performance relevant data, such as execution time, operation counts, and memory or network traffic to help application developers pinpoint and ultimately fix performance problems. There now exist a number of standardized tools and APIs to collect a wide range of performance data at the largest scale. However, this success has created a new challenge, as the resulting data is far too large and too complex to be analyzed in a straightforward manner. While there exist some tools for performance analysis and visualization, these are predominately restricted to simple plots of the raw data and rely virtually exclusively on the users to infer connections between measurements and the observed behavior and to draw conclusions. Unfortunately, as the number of cores increases, this approach does not scale. The raw data is typically rather abstract, low-level, and unintuitive and it is difficult to understand within the context of the highly complex interaction of an application with the middle- and system software and the underlying hardware. For this reason, new automatic and more scalable analysis approaches must be developed to allow application developers to intuitively understand the multiple, interdependent effects that their algorithmic choices have on the resulting performance.
Following classical visualization mantra, the natural first step towards automatic analysis is to display an overview of the collected data to provide some insight into general trends. This helps both application developers and performance experts to form new hypotheses on potential causes of and solutions to performance problems. Furthermore, intuitive visualizations are highly effective in conveying the results of any analysis and thus are a valuable tool throughout the entire process. Unfortunately, visualizing performance data has proven challenging as the information is highly abstract, non-spatial, and often categorical. While some early attempts at including more advanced visualizations in performance tools have been proposed, these are rudimentary at best and have not found widespread adoption.
At the same time there exists a vibrant community in the area of information visualization and lately visual analytics that is exclusively aimed at developing techniques to visualize, illustrate, and analyze complex, non-spatial data. In particular, there exists a large body of work on general design principles of visualization tools, color spaces, and user interfaces as well as a wide array of common techniques that tackle a broad range of applications. The Dagstuhl Perspectives Workshop, for the first time, gathered leading experts from both the fields of visualization and performance analysis for joint discussions on existing solutions, open problems, and the potential opportunities for future collaborations.
The week started with a number of keynote sessions from well-known authorities in each area to introduce the necessary background and form a common baseline for later discussions. It soon became apparent that there exists a significant overlap in the common tasks and challenges in performance analysis and the abstract problem definitions and concepts common in visualization research. Subsequently, the workshop continued with short talks focusing on various more specific aspects of either existing challenges or potential solutions interspersed with increasingly longer group discussions. Theses extensive, inclusive, and in-depth exchanges ultimately shaped the second half of the workshop and in this form were only made possible through Dagstuhl's unique collaborative and discussion stimulating environment.
Ultimately, the workshop has started a number of collaborations and research projects between previously disparate fields with the potential of significant impact in both areas. Furthermore, the participants distilled the open challenges into three high-level recommendations: First, joined funding for the various open research questions. Second, support to build and foster a new community on the border of visualization and performance analysis. And Third, the need to better integrate the anticipated results into the entire lifecycle of a massively parallel application from design to optimization and production.
Creative Commons BY 3.0 Unported license
Peer-Timo Bremer and Bernd Mohr and Valerio Pascucci and Martin Schulz
- Computer Graphics / Computer Vision
- Large scale data presentation and analysis
- Exascale class machine optimization
- Performance data analysis and root cause detection
- High dimensional data representation