Page 305 - The Combined Finite-Discrete Element Method

288 COMPUTATIONAL ASPECTS

is a separate PC with its own CPU and RAM, and most often hard disk. In some distributed
computing applications, different computer types can be used, even different operating
system. For combined finite-discrete element simulations this is not desirable, and the
best solution is probably a cluster of identical PCs configured in such a way that mapping
from sub-domains onto the PC cluster is relatively simple. In this way, communications
between PCs are kept to a minimum, and the amount of time each CPU spends actually
processing the job assigned to it is maximised.

9.5.3 Grid computing

Grid computing can be defined as a massive integration of computer systems available
through a network to offer performance unattainable by any single machine. Grid com-
puting enables the virtualisation of computing resources distributed over a grid. These
include processing, network bandwidth and storage capacity, which are used to create
a single virtual image of the system, For grid-based combined finite-discrete element
simulations, domain subdivision is obviously one of the ways of exploiting the grid. An
alternative and more feasible way is to use the grid as a virtual experimentation lab.
In that way, instead of using the grid as an alternative to massively parallel or cluster-
based simulations, it is used to complement them. Researchers located at various locations
around the globe can in essence simultaneously work together on the same problem. An
example of such a class of problems is parametric studies, where it may not be necessary
to repeat large scale computations –the results can be parameterised instead, and made
available to both research and industry. Such combined finite-discrete element projects
would rely on individual users harnessing the unused processing power and coordinating
the work towards a common goal. Thus, both money and resources are saved, the project
is speeded up and cooperation among individual researchers is brought to a new level.
A particular combined finite-discrete element simulation is nothing more but a numeri-
cal experiment. Much like a given physical experiment in the lab, the results of such a
numerical experiment should be independently verified and/or validated. Verification in
essence is about confirming that everything was done properly and that the results are not
a consequence of a coding error in the program. Validation is the next step, matching the
results of a numerical experiment to the results of a physical experiment. To grid-enable
such processes, in the era when numerical experimentation is to some extent replacing or
complementing physical experimentation, it is not enough to communicate results through
journals. The space available for journal publications is too small to record all the details
of a numerical experiment. While a physical experiment can very often be recorded on
a couple of pages of written text, a numerical experiment may involve a large amount
of input data, problem parameters, methods used, algorithms employed, implementation
details, etc. The only way to solve the problem is to have both hardware and software
tools in virtual form, enabling easy abstraction of such problems and also easy access to
problem data through a virtualised distributed database.
In summary, various options are available to address very large scale and grand chal-
lenge discontinua problems. However, it appears that, due to a need for communication
between processing units, there is a limit to what speedup can be achieved using any of
the above listed options. Existing parallel, distributed and grid computing options are able
to achieve better CPU and RAM performance, thus increasing the size of the problem