Boost logo

Glas :

RE: [glas] multilevel block-matrices

From: Edwards, Harold C (hcedwar_at_[hidden])
Date: 2005-03-17 11:00:47


The abstractions for vector spaces should be rich/flexible enough to
capture the needs of parallelism. I've found it straight forward to
apply the concept of vector spaces and subspaces - e.g. the partitioning
of storage or computational responsibilities among processors is defined
by a set of subspaces, one per processor, of the vector space.

Supporting subspaces and sets of subspaces has many other applications,
e.g. defining multi-level block partitioning, defining the resizing of
vectors as a re-assignment to a different subspace.

--Carter
  

> -----Original Message-----
> From: glas-bounces_at_[hidden]
> [mailto:glas-bounces_at_[hidden]] On Behalf Of Karl Meerbergen
> Sent: Thursday, March 17, 2005 8:36 AM
> To: Generic Linear Algorithm Software
> Subject: Re: [glas] multilevel block-matrices
>
> On Friday 11 March 2005 17:23, plagne.laurent_at_[hidden] wrote:
> > Hi all,
> >
>
>
> > I wonder if this kind of feature :
> > - multi-level blocked matrices.
> > - optional storage policy (virtual/actual) for matrices.
> > can be of more general interest and should be considered in the GLAS
> > project ?
>
> This is of interest. At this stage, however, we are still
> defining the basic
> concepts.
>
> >
> > I also wonder about the best level for parallelism consideration :
> > * Computer science level : e.g. STAPL
> > * Linear algebra level (GLASS level ;-))
> > * Higher Level : e.g. FEM solvers.
>
> Different levels may have to be considered. Parallelism was
> one of the options
> that was mentioned in the early requirements for GLAS. It all
> depends where
> the parallelism shows up in the application code: in a FE
> code, it can appear
> on the domain level, but it can also appear in the linear solvers.
>
> At this stage, I would like to see the basic concepts
> defined. There appears
> to be a large interest to define a VectorSpace concept, so
> let us do that
> first and see what needs to be done next.
>
>
> Karl