Design and Implementation of Computational Steering for Parallel Scientific Applications

Computational steering facilities allow users to interactively monitor and control the progress of their applications. They also allow application cross-steering, where results from one application are used to feed and guide another application. This situation typically occurs in multidisciplinary design optimizations and other complex meta-applications. In this paper, we discuss the computational steering facilities provided by the Distributed Resource Management System (DRMS). The DRMS programming model is based on user-defined schedulable and observable points (SOPs). At an SOP, controllable and observable variables can be examined and modified, data distributions can be changed, resources can be allocated, and a snapshot of the application can be taken. What-if analysis of applications is well supported by this environment. We describe the implementation and evaluate the performance of array section streaming operations in DRMS. These operations are important in the context of data visualization and cross-steering, and our results show that they can be performed efficiently on the IBM SP2.