In the *task farming approach*, each node runs its own copy of a
neural simulation (generally a detailed single-cell model). Each node
and, therefore, each simulation runs totally independently of all other
nodes. This method is particularly suited to examining large parameter
spaces. In many of our applications, there are a wide variety of free
parameters (i.e., those not defined experimentally). By using the task
farming approach on these supercomputer-class machines, we can range
widely across this huge parameter space looking for combinations which
give biologically realistic results [Bhalla:93a] (i.e., similar to
those measured experimentally). This allows us to make predictions for
the future experimental measurement of these free parameters. It is
also possible to run the same model many times in order to build up
statistically significant summaries of the overall model behavior. The
task farming approach is inherently parallel (zero communication between
nodes and, therefore, linear scaling of computation with number of nodes
available) and as such it is one of the most efficient programming
styles available on any parallel machine (i.e., it allows the full
utilization of the available computational power of the machine). This
approach allows modelling in a single overnight run what would otherwise
take a full year of nonstop computation on previously available
computing platforms.

Wed Mar 1 10:19:35 EST 1995