The Workshop on Automatic Performance Analysis (WAPA 2005, Dagstuhl Seminar 05501), held December 13-16, 2005, brought together performance researchers, developers, and practitioners with the goal of better understanding the methods, techniques, and tools that are needed for the automation of performance analysis for high performance computing.

High Performance Computing is an crucial component of current and future advances in science and engineering. These advances depend on simulations that run on massively parallel systems to obtain results within a reasonable response time. HPC are becoming commonplace, but exist in a wide variety of forms, and the forms dictate what kind of computation will running efficiently. Most current HPC systems are clustered SMP architectures with a private address space for each node. A few large-scale shared memory systems exist with a NUMA characteristic. Architectural aspects, like the memory hierarchy within a node, the latency and bandwidth of the communication network, and the connection of the node to the network are exposed to the programmer. Their individual complexity and mutual dependencies make the process of iterative program tuning, combining performance measurement, analysis, and optimization, an essential technique for high performance.

The workshop approached this problem by addressing techniques for automating the performance analysis process. Automatic performance analysis will: enable analysis of large application runs with a different behavior than scaled-down versions, allow triggering on-the-fly optimizations, and free the application expert from technical details of the analysis process. New advances in performance analysis automation are timely and achievable, and are the first step towards a fully automated tuning process.

WAPA 2005 builds on the successful Dagstuhl Seminar 02341 on Performance Analysis and Distributed Computing held in August 2002. The workshop is also a continuation of the European Working Group on Automatic Performance Analysis which ended in July, 2005. The experiences from those two activities played an important role in shaping the WAPA focus.

The WAPA program allowed a variety of opportunities for interaction among the workshop participants. Many leaders in the HPC performance community were able to attendWAPA and contribute to the technical content and discussion. As outlined in the attached agenda, the workshop was loosely organized around four themes:

• Performance analysis for large-scale parallel systems
• Automated performance analysis and diagnosis
• Automated performance tuning and performance prediction
• Performance tools and technology

Each theme session consisted of a set of technical talks. In addition, we were fortunate to have two perspective talks by John Levesque (Cray Inc.) and Phil Roth (Oak Ridge National Laboratory, on behalf of Jeff Vetter). Complementing the technical presentations, two panels provided an interesting perspective on the state of performance analysis:

• What works? What’s missing? What’s all this about productivity?
• Wizard Tales and Santa’s Wish List

To complete the program, there was the opportunity for performance tool demonstrations. Eight tools were presented covering the techniques for HPC distributed memory analysis , scalable tracing, open performance technology, performance databases, and performance data mining.To complete the program, there was the opportunity for performance tool demonstrations. Eight tools were presented covering the techniques for HPC distributed memory analysis , scalable tracing, open performance technology, performance databases, and performance data mining.

The goals of WAPA were to increase the exchange of ideas among the tool developers, to transfer knowledge on existing and planned automatic techniques, to engage people supporting application development, and to start a dialog between researchers, developers, and users of automatic performance analysis methods and tools. By all accounts, theWAPA meeting was a tremendous success. Many participants commented on the high quality of the technical talks and the fruitful discussions they had during the week. And of course, everyone enjoyed our excursion and dinner in the historic city of Trier.

It is the energy and meaningful involvement of each participant that is most responsible for the positive WAPA experience. The organizers would like to thank everyone who came to Schloss Dagstuhl and we look forward to when our paths cross again in the future.