Balancing consistency and lag in transaction-based computational steering

Computational steering, the interactive adjustment of application parameters and allocation of resources, is a promising technique for higher-productivity simulations, finer-grained optimization of dynamically varying algorithms, and greater understanding of program behavior, data sets and solution spaces. Typically, these "steerable" programs have been implemented as parallel or distributed programs. Tools for computational steering must provide monitoring, visualization, and interaction facilities, must address issues related to consistency, latency, and scalability at each of these phases, and must consider perturbation effects. In this paper we describe transaction-based components for a computational steering system and present an approach that guarantees consistent monitoring and displays, supports scalable monitoring, and provides the user with the ability to adjust the tradeoffs among lag, consistency and perturbation.