Compute-Intensive Simulations for Cellular Models - Parallel Computing for Bioinformatics and Computational Biology - page 99

Figure 4.4 MathML Interface for performing stochastic chemistry in a Grid environment.
data movement (in this case through the chemical signaling), and is also difficult to
“scale” as the number of processors increases.
An easier alternative is to perform domain decomposition at the data level. This is
often called data parallelizm and involves different processors performing the same
function but on different sets of data. The data need to be spread among the processors
in a manner that minimizes the data transfer costs. Load-balancing must also be taken
into account; it is inefficient to have some processors idle while others are busy.
Data decomposition can be either explicit (where the programmer directs how the
data are placed on the processors) or implicit (which is where a compiler directive
may distribute some array data among the processors). In particular, the data may be
distributed cyclically by rows or columns, or subblocks of data could be distributed.
In some cases, data may be replicated locally to minimize communication costs.