Lack of effective performance evaluation and prediction environments is a major barrier to the broader use of high performance computing. Conventional performance environments are based on profiling and event instrumentation. For a system with p processors, there are 2^p-p-1 potential parallel sub-task interactions. As parallel systems scale to hundreds of nodes and beyond, the conventional execution profiling approach becomes problematic. Moreover, there are many ways to parallelize an application, and the relative performance of different parallelizations vary with problem size and system ensemble size. The traditional profile analysis approach does not meet today's needs, and is far from adequate for the emerging massively parallel peta-computers and world-wide virtual machines.
The objective of the proposed research is to explore the relevant issues in designing and developing an integrated, concrete and robust SCALability Analyzer (SCALA) system for performance modeling and prediction of high performance computing programs. In contrast to existing performance tools, the program performance model generated by SCALA is based on scalability analysis. The SCALA system is enabled by our recent success in scalability study and in integrating scalability analysis into parallel restructuring systems to predict the performance crossing point of different parallel implementations a utomatically.
SCALA assumes the availability of modern compiler technology, adopts the synthetic-perturbations screening and other statistical technologies, and has the support of knowledge database. These technologies, together with a new approach of scalability, enable SCALA to provide the user with a higher and more intuitive level of performance analysis. SCALA is designed to explore the validity of these new techniques, and use them collectively to bring performance analysis environment to the most advanced level. SCALA does not compete with existing or on-going performance environments.Instead, SCALA can be used in connection with existing performance monitoring and analysis tools and will be accessible via the Web.
SCALA provides the user with analytical models of the observed performance and its scalability. No current environment, with this level of automation does that.
The development of SCALA is a joint project of the SCS group and the Institute for Software Technology and Parallel System, directed by Professor Hans Zima, at University of Vienna. Dr. Mario Pantano, while serving as the coordinator of the ICE project, is directly responsible for the development of SCALA.
Back to SCS Home Page