CS 240A HW0
Ting Lei
The application I have investigated is "Elastic Properties of Concrete". It is a parallel computing project of NIST's
Information Technology Laboratory.
The purpose of the project is to use parallel computing methods to find
out the elastic moduli of concrete (and other materials). An elastic
modulus is the ratio between the stress applied on a substance and
the tendency of the substence to be deformed (strain). Elastic moduli
of concrete characterizes many of its non-fracture-related mechanic
properties, and is in many highrise buildings, more important than the
stiffness of the material. The project predicts the overall elastic moduli of concrete or any appropriate material given the 3-D image of its micro-structure and individual phase
elastic moduli.
The parallel platform that they used for reporting results is an
8-CPUs SGI Origin 2000. An web search will show that this machine
is of a symmetric multiprocessing (SMP) shared memory architecture.
Given the number of CPUs there, I would not think it is a major
supercomputer. Their code is written with MPI and Fortran 90, and it
can reportedly be run on distributed memory system too.
In terms of performance, they reported a speed-up of 18 for a 3-D image
of 3000^3 voxels as compared to nominal running time for the serial
algorithm. As there are only 8 CPUs, the effiency would be more than
100%. Therefore I suspect there must be unfair assumptions made about
the "nominal" serial program. However, I think they do achieve their
goal for processing problems of larger sizes. It is obvious that if the
code is 3-D image resolution is doubled on each dimension, the number
of voxels will be 8 times more. Using their parallel algorithm allows
processing larger pieces of the material or the same piece with finer
resolutions.
I think their method will scale well to large problems because their
method looks similar to finite element and there is obviously spatial
locality in the problem and this can probably be exploited to decompose
the computation tasks to some degree.