linpack/hpl

[Ray Jones]

Greg Lindahl <lindahl at conservativecomputer.com> writes:

> On Tue, Sep 25, 2001 at 08:22:41AM -0700, Shane Canon wrote:
> 
> > We are adjusting the problem
> > size and block size primarily.
> 
> You need to set the problem size as large as possible.  If it were too
> small, you'd see exactly what you report, namely that above a certain
> # of nodes, it begins to suck.

Linpack is also be very sensitive to P and Q, the dimensions of the
virtual grid that your nodes are organized into for the computation.
In my experience (limited to 1 cluster), more square grids are best.

Ray