Your comments about my abstract titled "Parallel Optimization in Engineering" are welcome
Henry Chan
The abstract of my paper titled "Parallel Optimization in Engineering"
has been accepted by the Workshop on Nonlinear Model Based Control -
Software and Applications (NMPC-SOFAP, 2007) dated April 19-20, 2007
located at Loughborough, United Kingdom.
I would like to have some suggestions and comments from all of you
about my abstract and/or presentation.
Your valuable points and constructive criticisms are all welcome.
Thanks for your kind help and comments in advance.
Best regards,
Henry Kar Ming Chan
_____________________________________________________________________________
Title : Parallel Optimization in Engineering
Author : Henry Kar Ming Chan
For the parallel optimization in computational science and
engineering, it is required to solve the large scales non-linear
optimization problems in parallel processing mode. OpenMP supports
multi-platform shared memory multi-processing architecture and is an
industrial standard Application Programming Interface(API) for forking
and joining program and sub-program tasks. From the product of
optimizers, Knitro is an advanced solver for non-linear optimization
problems, handling bound constraints, non-linear equalities and
inequalities (both convex and non-convex), and complementarity
constraints. This is because Knitro mainly has the state-of-the-art
features using the algorithmic options, such as interior point methods
and active set method, which can be switched between one another
automatically or manually by changing the parameters settings. Knitro
can handle the inequality constraints by an interior point algorithm
and direct solution of the barrier sub-problems. In large scales non-
linear problems, Knitro can handle the inequality constraints by an
interior point algorithm and solution of the barrier sub-problems by
conjugate gradient iterations. For detecting infeasibility, Knitro can
handle the inequality constraints by an active set algorithm, which is
especially beneficial when a good initial point is available and when
solving a sequence of related problems. In some cases, the initial
guess provided by the user and all subsequent iterates also satisfy
these inequality constraints, then a feasible mode can be set in the
parameters list. A simple test scenario is used as test cases.
Different active set methods are tested in parallel processing using
OpenMP section or task parallelism. Within each active set methods,
multi-start features are used to randomly select the initial points so
as to find the global optimization points if possible. Parallel Cplex
for linear programming solver and Intel MKL library for vector and
matrix processing are used with OpenMP directives. Gradient and
Hessian vectors for directional and partial derivatives are generated
with OpenMP features through Automatic Differentiation. Finally,
vectorization of data, pre-fetching of data and inter-procedural
optimization for parallel OpenMP programs are necessary to be
considered during program compilation to achieve the best
performance.
____________________________________________________________________________