Taking modelling to a different plane
Background
BAE Systems is a global company engaged in
the development, delivery and support of advanced defence and
aerospace systems in the air, on land and at sea.
They are the largest European defence company with annual sales
exceeding £12 billion ($22 billion) and 88,000 employees.
The problem
Computer modelling has become an indispensable tool in modern
engineering design. The aerospace industry makes extensive use of
computational fluid dynamics (CFD) simulations in the design of new
aircraft. BAE SYSTEMS have an in-house CFD code which they use to
model the aerodynamics of military aeroplanes. Accurate prediction
of the airflows around the whole aircraft is important in
determining the heat and noise characteristics in operation, both
of which impact on running costs. However, whole aircraft
simulations were impossible to achieve using single processor
hardware. Ground breaking code development was required to allow
this.
The challenge
Solving equations of fluid flow is non-trivial even for simple
geometries. Solving them exactly for complicated geometries such as
a whole aircraft is impossible. Best approximations are approached
by constructing 3D finite element grids around the surface of
interest. Grid mesh size varies depending on the rate of change of
the property being modelled. Grids of over 10 million nodes are
possible; however they can only be run in reasonable timescales
using parallel computing techniques. The challenge was to convert
the BAE SYSTEMS in-house code to operate in parallel mode in order
to enable accurate whole aircraft simulations to be performed in
acceptable timescales.
The solution
The development of parallel computing techniques is based on
partitioning the grid into roughly equal sections and then
distributing each section to a separate processor. Inter-processor
communication is used to update points on the surface of each
sub-grid. In practice, a three stage process is used involving
coarsening, re-partitioning and refinement. The method developed at
Daresbury is self adapting, inserting extra grid points as
necessary where the solution varies rapidly. This can lead to load
imbalance among the several processors and so a load sharing
mechanism is also used. Using a 128 processor distributed memory
architecture this approach succeeded in delivering whole aircraft
simulation involving over half a million nodes and, in addition,
the solution was achieved 80 times faster than it would have been
on a single processor of equivalent capacity.
The benefits
The customer has gained an ability to perform computer
simulations that were previously not achievable.
Whole aircraft simulations can now be performed in a fraction of
the time taken previously, allowing design improvements leading to
product optimisation and significant savings in both time and
money.
The customer has effected a permanent improvement, keeping it at
the forefront of aerospace technology.