Parallel Simulation of Multiprocessor Systems Using Reverse Execution

This project is exploring the use of parallel discrete event simulation techniques to perform instructionlevel simulations of large-scale multiprocessor systems. The large memory and computational requirements of such simulations and the inherent parallelism of the problem domain suggest the use of parallel simulation techniques. However, a critical challenge lies in ensuring proper synchronization of the parallel computation. Optimistic synchronization techniques such as Jefferson’s Time Warp algorithm offer promise to addressing this issue, but have large memory requirements because past state information must be saved to allow one to roll back the computation. We propose the use of reverse execution techniques to greatly reduce the memory required to perform optimistic parallel simulations by re-computing past state information as needed, rather than using traditional state saving mechanisms. This paper describes preliminary work to assess the feasibility of this approach. Specifically, key aspects of the approach are described, as well as initial thoughts on how the generation of reverse execution code can be automated.