TetraPower 970MP (CANCELED) Project
Personal Supercomputingin category Applications & Software
proposed by kschwebke on 27th May 2006
Project Proposal
Personal SupercomputingMany fields of research are supported by applications needing
decent computing power. In the past, the only option was batch
or online access to a remote supercomputer shared with other
users and therefore additional coordination overhead.
Now researches have another option: personal supercomputing.
The exclusive access to a high performance computer accelerates
and eases the development of research applications
(e.g. debugging or benchmark).
The Project
In this project a research software package for the Generalized
Finite Element Method (GFEM) will be ported to the PPC platform.
The main demand for processing power arises from solving a special,
large and semidefinite linear equation system. Popular solvers,
e.g. multi-grid, can only solve for definite systems.
The algorithm used allows hybrid parallelization:
- The problem domain is decomposed into weakly coupled
sub-domains, which can be solved with low synchronization
overhead. The MPI (message passing interface) API is
applied here. - The sub-domains may be subject to a further local
parallelization exploiting shared memory multiprocessing
(pthread API) or vectorized matrix-vector multiply operations.
The package already runs on AMD Athlon or Intel IA-32/Linux personal computers
and SGI Origin/Irix NUMA supercomputers. The speedup on SGI Origin platform
comparing 1 to 10 processors is in the range of 6...8.
On porting to PPC platform the main effort will be implementing
vectorized and shared memory multiprocessing (SMP).
To enable the building of multi-node clusters the message passing paradigm
will be covered, too. This will not be the main focus, however.
More on GFEM: http://www.bauv.unibw-muenchen.de/bauv1/oc/html/forschung/gfem/
Project Blog Entries
