FNM

Abstract

As multicore computer systems become increasingly complex, parallel simulation is becoming an important tool for exploring design space and evaluating design tradeoffs. The key to the success of parallel simulation is the ability to maintain a high degree of parallelism under synchronization constraints. In this article, an enhanced Null-message algorithm called FNM is presented that uses domain-specific knowledge to improve the performance of the basic Null-message algorithm. Based on their runtime states, the components of the simulation model can make a conservative forecast of future interprocess events. The forecast information is carried in the enhanced Null-messages, and, by combining the forecast from both sides of an interprocess link, FNM can achieve a dynamic system lookahead that is much greater than what the static system structure provides. This improved lookahead allows better exploitation of the simulation model's inherent parallelism and leads to better performance. Compared with the basic Null-message algorithm, FNM greatly reduces the amount of Null-messages and improves parallel simulation performance as a result, while at the same time it guarantees simulation correctness as the basic Null-message algorithm does. In tests on cycle-level models with up to 128 cores, FNM shows good scalability and proves to be an effective method.