Fine-tuning optimistic simulation

Abstract

The chief characteristic that differs markedly between parallel simulation techniques is how they manage process interaction. The linear control models for online configuration of the simulation presented in this article are different from traditional control theory. The reason is because data sampling and parameter adjustment are intrusive; these operations contend for processor cycles that can be utilized for useful computation. In addition, this control process is by necessity imprecise. First, the changes of the local logical process (LP) configurations will have secondary and tertiary effects that are reflected back to the LP. These effects are difficult to quantify. Second, because data sampling and parameter adjustment are intrusive, it is usually not feasible to implement the most accurate control system if it is too complex. The online configuration model was used to adapt the behavior of different facets of the simulation. Because of the imprecise nature of the control systems, it is nearly impossible to find the optimal settings for them. We relied on empirical data to guide the design of the control systems. In particular, control strategies for dynamic adjustment of the checkpoint interval and the cancellation strategy (lazy or aggressive) are presented. Performance results for the control strategies demonstrate that dynamic schemes can considerably improve the performance of a time warp-based parallel discrete-event simulator. In experiments conducted on several application domains, the performance of the time warp simulator was improved by dynamically limiting the overheads of the time warp algorithm.