Biomedical Engineering Reference
In-Depth Information
computation does not use stale values. Therefore, naive parallelizations, where
the activity of each processor is restricted to a predefined subdomain of the
total lattice, will increase the frequency of interprocessor communication for
synchronization and, eventually, the waiting time of each node will be much
greater than its calculating time. An interesting attempt to overcome this
issue has been made with the checkerboard algorithm, presented in [81]. Its
authors have used an improved data structure to describe simulated individ-
uals and have further decomposed the subdomain assigned to each node into
smaller subgrids (i.e., sublattices), chosen so that the corresponding ones on
different processors do not interact. In this way, an update in one sublattice
affects only the nearest-neighbor sublattices. Each node is therefore able to
determine the spin flips affecting neighbor nodes, accumulate them, and pass
them synchronously. In this case, the speed gain increases with the size of
the subgrids per processor and decreases with the interaction range. Such a
basic checkerboard parallelization can also implement rejection-free or RW
methods, by using either equivalent or master-slave computations [20].
Search WWH ::
Custom Search