Building a Community Infrastructure For Scalable On-Line Performance Analysis

Peta-scale computing environments pose significant challenges for both system and application developers, and addressing them will require more than simply scaling up existing tera-scale solutions. To exploit the capabilities of next-generation systems, it will be essential for developers to gain a detailed understanding of interactions between application components, system software, and hardware. Performance analysis tools will play an important role in gaining this understanding but existing monolithic tools with fixed feature sets will not suffice. Instead, this project will design, implement, and evaluate a general, flexible tool infrastructure supporting the construction of on-line performance tools as "pipelines" of high-quality tool building blocks. These tool building blocks provide common performance tool functionality, and are designed for scalability, lightweight data acquisition and analysis, and interoperability. For this project, we build on Open|SpeedShop, a modular and extensible open source performance analysis tool set. This project is a collaboration between researchers at Krell Institute, Lawrence Livermore National Laboratory, Oak Ridge National Laboratory (ORNL), the University of Wisconsin-Madison, and the University of Maryland.