Supercomputers in Grids - Cloud, Grid and High Performance Computing: Emerging Applications - page 6

logical drivers of the main stream IT-technology
development.

tion complex scenario crash,) by nature cannot be
made embarrassingly parallel and any distributed
Grid solution has so far failed on them.
The time to solution: Most of the large scale
problems mentioned above actually can run on
smaller systems. However, on such smaller sys-
tems their solution may take weeks or even months.
For any practical purpose such simulations would
make little sense. The Grid is hence unable to
provide scientists with a tool for these simulation
experiments if it aims to replace supercomputers
by a large amount of distributed systems.

Grids Cannot Replace
Supercomputers

Sometimes the Grid is considered to be a replace-
ment for supercomputers. The reasoning behind
this idea is that the Grid provides such a massive
amount of CPU cycles that any problem can eas-
ily be solved “on the Grid”. The basic concept
for such reasoning is the premise that a given
problem can be described in terms of required
CPU cycles needed. On the other hand, any given
Grid configuration can be described in terms of
CPU cycles provided. If one can match compute
demand and compute supply, the problem is as-
sumed to be solved.
This is, however, a deeply flawed view of
supercomputing. The purpose of a supercomputer
is to provide the necessary speed of calculation to
solve a complex problem in an acceptable time.
Only when being able to focus a huge resource
on a single problem can we achieve this goal. So,
two aspects are important here.
The size of a problem: We know of a number
of problems that we call large which can actually
be split into several small problems. For such em-
barrassingly parallel problems the Grid typically
is a very good solution. A number of approaches
have been developed among which Berkeley Open
Infrastructure for Network Computing (BOINC
2008) and the World Community Grid (2008) are
the most interesting ones. Both provide access
to distributed resources for problems that can be
split into very small junks of work. These small
problems are sent out to a mass of computers
(virtually every PC can be used). Doing this,
the systems are able to tap into the Petaflops of
performance available across the globe in an ac-
cumulation of small computers. However, there
are other large scale problems that cannot be split
into independent smaller parts. These truly large
scale problems (high resolution CFD, high resolu-

THE ROLE OF
SUPERCOMPUTERS IN GRIDS

The Grid has often been compared to the power grid
(Chetty and Buyya, 2002). It actually is useful to
look at the power grid as an analogy for any Grid
to be set up. Power Grids are characterized by:

• A core of view production facilities pro-

viding a differing level of performance
much higher than the need of any single
user. Small facilities complement the over-
all power Grid.

• A very large number of users that typically

require a very small level of performance
compared to the production capacity of the
providers.

• A standardized way of bringing suppliers

and users together.

• A loosely coordinated operation of suppli-

ers across large geographic areas.

• Breakdowns of the overall system if coor-

dination is too loose or if single points of
failure are hit.

• Special arrangements for users requiring a

very high level of performance on a per-
manent basis. These are typically large
scale production facilities like aluminum
production.