Falcon { Toward Interactive Parallel Programs :

This paper focuses on the opportunities and costs of on-line steering as applied to a substantial parallel application. We demonstrate potential performance improvements through the use of the Falcon system, an experimental system for the on-line monitoring and steering of parallel programs. The visual presentation of program output along with animated displays of program performance information via Falcon's monitoring system enables the on-line capture, analysis, and display of program information required for program steering. Falcon also provides the mechanisms for the manipulations of program state that accomplish this online steering. 1 Interactive parallel programs Future parallel applications will run on a variety of parallel machines linked with high performance networks. The primary purpose of our research is to exploit these modern systems' capabilities to ooer human-interactive interfaces that can execute simultaneously with a parallel application's computational and storage tasks. Speciically, we are exploring the potential for increases in performance and function-ality gained by the on-line interaction of end users with their supercomputer applications on single and on networked parallel machines. Unlike the eeorts of the scientiic visualization community which are often targeted towards interaction with the data produced by parallel programs, we are concerned with interaction with the programs themselves. Such on-line interactions have already been identiied as highly desirable by both end users and program implementors, and these interactions are collectively referred to as interactive program steering. The program steering addressed by our work encompasses rapid, on-line program changes that require built-in, custom monitoring and on-line algorithms determining and enacting such changes (e.g., in research addressing conngurable operating systemssMS93]). It also includes gradual changes to long-running real-time programssBS91] and scientiic applications able to interact at human speeds via user interfaces, such as the application presented in this paper. Regardless of the speed and nature of interaction, program steering is based on the on-line capture of information deening current program and connguration stateeBru91, MW91, OSS93], and it assumes that human users or algorithms inspect such information and manipulate it to make steering decisions. This paper focuses on the opportunities and costs of on-line steering as applied to one substantial parallel application, a molecular dynamics simulation (MD) used by physicists for exploration of elementary properties of lubricants. We demonstrate the potential of improving performance by on-line steering with experimental results attained on a 64-node KSR-1 su-percomputer. An experimental system for the on-line monitoring and steering of parallel programs { the Falcon …