Experiences using Computational Steering on Existing Scienti c Applications

Computational steering remains an impressive yet challenging opportunity for computational scientists. In this paper, we examine our practical experiences when we used computational steering on four existing scienti c applications. In particular, this paper enumerates the concrete steps required for speci c computational steering on each individual application. We then provide an evaluation of steering e ects on application performance, and also present necessary modi cations to the application. Finally, we provide a summary of common issues distilled from these experiences. 1 Computational steering Simulations are playing an increasingly critical role in all areas of science and engineering. As the uses of these simulations expand, the demand grows for high performance computing of increasing power, exibility, and utility. Interactive computational steering is one way to increase the utility of high performance simulations for scientists because it allows them to drive the scienti c discovery process and interact with their data. They can interpret what is happening to data during simulations and steer calculations in close-to-real-time: they can change parameters, data-sets, resolution, and representation; then, see the e ects. In this work, we explore some of the challenges for creating an eÆcient software infrastructure for computational steering. Speci cally, we consider two important requirements for interactive computational steering: steering latency and application perturbation. High latency results in poor decision quality and low steering frequency. High application perturbation con icts with the original intent of high performance applications. In addition, our approach assumes three basic constraints. First, we attempt to minimize the changes to the target application and system. Literally, we try to minimize the number of changes to the source code and to preserve the application's original execution environment. Second, we strive to minimize the di erences in the infrastructure across applications. Our objective is to develop a general set of mechanisms to change applications even though applications exhibit a variety of designs and behaviors (e.g., data decompositions versus functional decompositions). These mechanisms used in conjunction with speci c policies and metaphors provide application-speci c steering. Finally, we try A portion of this work was completed while Vetter was a PhD candidate at the Georgia Institute of Technology. The Georgia Tech work was funded in part by a NASA Graduate Student Researchers Program Fellowship for Vetter, by NSF equipment grants CDA-9501637, CDA-9422033, and ECS-9411846, and by Los Alamos National Lab. At Illinois, this work is funded in part by the Defense Advanced Research Projects Agency under DARPA contracts DABT63-94-C0049 (SIO Initiative), F30602-96-C-0161, and DABT63-96C-0027, by the National Science Foundation under grants NSF CDA 94-01124 and ASC 97-20202, and by the Department of Energy under contracts DOE B-341494, W-7405-ENG-48, and 1-B-333164. Department of Computer Science, University of Illinois, Urbana, IL.