ARNG:Accelerating discrete event simulation RNGs by asynchronous random number generation

Functional decomposition, also called functional parallelism, was an approach to introduce inner parallelism into discrete event simulators in the 1980s, in order to accelerate simulation experiments. Due to technical restrictions at that time, it did not gain widespread acceptance. This paper introduces modifications of the approach, considering today's technical possibilities. A focus is set on asynchronous execution of simulation infrastructure by thread pools. Applying the concepts of asynchronous functional parallelism to random number generation leads to our proposal for asynchronous random number generation (ARNG), which has been implemented into the discrete event simulation framework DESMOJ. We describe its implementation and report on results of experiments which were conducted to assess the performance potential. The impact of different parameters is analyzed and advice for parameterization is given. We found that the attainable acceleration factor of our ARNG implementation is limited within a range from circa 2 to 5 on an eight core machine, depending on the concrete distribution and the number of thread pool workers used. If two (otherwise idling) processor cores are employed, then random number generation can be accelerated by a factor of at least 1.74 in non-trivial simulation models, and by a factor of at least 3.00, utilizing four processor cores.