The cost of conservative synchronization in parallel discrete event simulations

Reviewer: Haim S. Gabrieli

Nicol makes a welcome contribution to current research on developing effective system simulations in multiprocessing environments. These simulations present a special challenge to their designers because events are generally time-dependent and concurrency must be maintained between the processors. The paper describes a protocol for parallelized discrete-event simulations that, based on the author's findings, will yield high processor utilization and low overhead. This conclusion is backed by analytical results as well as empirical measurements from several simulations run on the Intel iPSC distributed memory processor. The protocol is based on a sliding window concept defining lower and upper bounds within which events can be processed. Note, however, that the protocol will only work for a special class of simulation models that meet certain conditions. This paper will be of interest primarily to those engaged in designing and researching large-scale simulations in multi processing environments. A brief overview of recent work on the topic is provided. The distinction between the conservative and optimistic schools of thought is also discussed. If you have been exposed to simulation design and research, you are likely to find most of this paper easily readable. It provides a good blend of practical, down-to-earth discussion combined with a more rigorous mathematical analysis for those so inclined. It also includes a list of 32 references, most of which relate to work on parallelized simulations.