Feedback control in time warp synchronized parallel simulators

The time warp mechanism is one of the most important synchronization protocols for parallel simulation. However for most applications, the successful use of time warp requires the careful selection of time warp optimization parameters (e.g. cancellation strategies, state saving frequency, and so on). Unfortunately, the optimal setting for the simulation parameters may not hold across an application domain or even throughout the entire simulation lifetime of a single application. Consequently, several investigators have proposed the dynamic adjustment of simulation parameters over the lifetime of the simulation. The dynamic adjustment of simulation parameters requires careful design considerations. Many of these considerations are similar to the problems studied by traditional nonlinear and adaptive control theorists. These considerations include the need to implement an adjustment mechanism that converges to stable values and the need to understand how parameter adjustment affects (sampled) output values. However, the dynamic adjustment of simulation parameters occurs by using general CPU cycles and thus introduces overhead into the simulation. This overhead must be carefully controlled-too little processing and the control system may not respond correctly; too much processing and the overall simulation performance may deteriorate. We address the design of two distinct feedback control systems for dynamic parameter adjustment and describe the design considerations that were made in their construction.