Parallel implementations of individual-based models in biology: bulletin- and non-bulletin-board approaches

Particle-based models are simulations in which the discrete representation of physical phenomenon involves interacting particles. This paper studies the efficiency of two different methods of implementing these models on a network of UNIX workstations. Two data parallel methods of modeling particles are tested: bulletin-board and non-bulletin-board. In the former method, the programs communicate through a logically shared, associative memory called a bulletin-board. The simulated particles are distributed among the workstations dynamically as the processing load on the processors changes. In the latter method, the particles are divided amongst the networked workstations statically at load time. The simulated system is a collection of ants moving and foraging in a two-dimensional space. This paper analyzes and compares the execution times of both implementations for different combinations of particles and number of workstation, using speed-up, tuple granularity and communication cost as measures. Analysis shows that the bulletin-board method is better for particle-based simulations when the correct granularity is chosen.