[vtkusers] paralleliso isnt.

[John Shalf]

Stanley Dunn wrote:
> 
> On Mon, 20 Aug 2001, Berk Geveci wrote:
> 
> > When I ran this program in our cluster, I also observed that it scales
> > very poorly. I think there are two reasons for this:
> > 1) IO: In our setting, the data files are on a shared disk which is mounted
> > via NFS. All processes start reading from the same disk, sharing the same
> > network switch. To test whether this is the case, one would have to put a
> > local copy of the data files to each machine.
> > 2) Balance: This problem is not balanced AT ALL. Some of the processes
> > do not generate anything when contouring, some others generate a lot. The
> > situation gets worse as the number of processes is increased.
> 
> When I was working on this last year with Scott Klasky at PPPL, it did
> scale very poorly and I believe that this is due in part to
> marshalling/unmarshalling the data.  I assume that the
> marshalling/unmarshalling has not changed in the past year.
> Stan

I think Jim Ahrens pointed out to me at one time that the
default marshalling/unmarshalling is in ASCII, but there is
a method that can be used to force the vtkMPICommunicator to
use binary.  Otherwise you get throttled by the cost of
marshalling/unmarshalling the data (an incredibly slow
process even on gigahertz CPU's).

Anyways, this was from a conversation from more than a year
ago about code that hadn't made it to release yet.  I expect
its now part of the vtk CVS tree, but I haven't had time to
follow up on it.  If anyone finds this method for turning
off ASCII marshalling, then please report to the list.

-john