|
Distributed simulations
Distributed simulation involves separately and concurrently
computing the dynamics of several parts of a system that is subject to natural
conservation laws, then coupling the response fragments into one comprehensive
solution at each integration time step. The distributed simulation can be
self-consistent and satisfy the natural conservation laws only if its results
coincide with the results of a conventional simulation of the same system. In
conventional (single computer) time domain simulations, the dynamic equations of
the system are discretized so that the system is represented by a system of
algebraic equations. The algebraic system is solved to yield the state of the
system at each time step. In the case of distributed simulations, the system’s
fragments are solved independently using conventional techniques, and then the
individual solutions are coupled by employing iterative methods such as
Gauss-Seidel or Gauss-Jacobi.
Examples of distributed simulations that use multiprocessing
systems can be found in the literature. Here, the data is usually exchanged
between the computational streams via the internal bus of the computer. In
network distributed simulations, on the other hand, the network serves as the
data link between individual workstations.
Network distributed simulations extend the possibilities of
the VTB time domain simulator. A user can partition a system into several parts
depending on the functional characteristics of the models that define the
system. Concurrent execution of the computational load among several
workstations has the potential to boost the computational speed. Further
improvements in speed can result from applying the most efficient solver
(different numerical techniques) on each workstation, as appropriate to the
particular dynamic subsystem.
Network distributed simulations have been implemented in VTB
in the following way. Several instances of the VTB software are run on many
workstations, all of which are networked. The system to be solved is partitioned
so that each instance of the VTB simulates its own fragment of the system.
Interactions between fragments are performed by of special network models. These
models handle the exchange of data and couple the solutions of each fragment
into the comprehensive system solution at each integration time step.
