Modeling and Simulation - Bioinformatics Computing - page 321

f(x) and lowering the system temperature T , the maximum function value will be trapped somewhere
in the global minimum. A major issue in simulated annealing is the cooling schedule, because
abruptly decreasing T may trap the function at a local minima. However, extending the cooling
schedule may lengthen the time required for the algorithm to locate the global minimum
unacceptably. Typical runs of the Metropolis Algorithm involve 100 to 1,000 iterations.
Execution time is often a significant issue in bioinformatics applications because many problems are
multidimensional. That is, instead of simply locating the global minimum in one plane, the
minimization problem is often one best represented in n -dimensional space, as in Figure 9-10 .
Because thousands of iterations in each dimension may be required to determine the global
minimum, computational time becomes prohibitive with increasing n .
Figure 9-10. 3D View of a Minimization Problem. Solving for global
minimum in a multi-dimensional space is computationally expensive.
Hardware
Simulations, especially those involving tens of thousands of data points and relationships, such as
those dealing with protein structure prediction, are extremely hardware-intensive. Many simulations
are beyond the capabilities of all but the most powerful general-purpose desktop workstations
operating at over 1 GHz with dual CPUs and several GB of RAM—and even these systems may take
days of processing time per simulation. The most affordable general-purpose alternatives to
mainframe hardware are to create a Linux cluster of affordable, modest-power workstations. A
cluster of 20 or more workstations can provide the computational power approaching that of a
mainframe at a fraction of the cost.
Depending on the nature of the simulation, specialized hardware may be available to make some
modeling and simulation tenable on desktop systems. For example, there are graphics accelerator
cards to enhance the rendering of molecules and other 3D structures. Similarly, for neural
network-based simulations, there are cards designed to represent the individual nodes in hardware,
speeding the lengthy learning process by several orders of magnitude.