No subject

These results were run before we had support for shared memory for local
communication. So, on our quad processor nodes, if any of the mpi
processes needed to communicate to another on the same board the
communication would have to go through the pci bus, myrinet card, to the
switch and back. We now have the shared memory support for local
communication which cuts the zero length, half round-trip latency by
better than a third. 

Doug

On Wed, 18 Oct 2000, Andreas Boklund wrote:

> 
> I replied to Daniel but i might as well translate my reply to him to
> english and add a few things and send it to the list.
> 
> I have build 2 clusters with the purpose of running fluent on them.
> 
> The first one was
> 
>  FoxHole
>  20 * Pentium III 500 	128SDRAM
>  1 * Cisco Catalyst 1900
> 
> 
>  The second one is the one that im "fine tuning" right now.
>  
>  ?????
>  1 * dual Pentium III 866 256RDRAM
>  9 * dual Pentium III 800 256RDRAM
>  1 * Cisco Catalyst 2900
> 
> 
> Our calculations has a LOT of chemistry in them so i dont think its fair
> to draw any standard conclusions from them. 
> But maybe a few pointers:
> 
> They dont require much bandwith.
> Although each iteration i do with my new cluster sends approx 150-200MB of
> information /iteration. One iteration takes just over a minute to
> complete. This might sound like a lot but the switch is not an issue.
> 
> On very few computers it has pretty good scalability but somewhere around
> 10-12 processors the gain of adding another one is very small. (we need
> the time so we added a few more anyway).
> 
> This chart http://www.exjobb.udd.htu.se/2000/sv3/25/Image3.gif (is in
> swedish) but it showes the time it takes for FOXHOLE with different number
> of nodes to complete our problem, and it shows the time it takes for a
> Dell with 4*550 and 1GBSDRAM. 
> 
> Fluent is pretty RAM hungry although for the nodes i "might" have been
> able to use 128MB but im considering to go to 512 on the master.
> 
> 
> 
> MVH 
> //Andreas
> 
> 
> And if anyone know Swedish (i know that Daniel does) you can read our
> Bachelor thesis on:
> http://www.exjobb.udd.htu.se/2000/sv3/25/
> Unfortunately we had to cut away a lot of the technical stuff :(
> 
> 
> 
>  *********************************************************
>  *  Administator of Amy(studentserver) and Sfinx(Iris23) *
>  *                                                       *
>  *   Voice: 070-7294401                                  *
>  *   ICQ: 12030399                                       *
>  *   Email: andreas at htu.se, boklund at linux.nu             *
>  *                                                       *
>  *   That is how you find me, How do -I- find you ?      *
>  *********************************************************
>  
>  
>  
> 
> 
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf
> 

-- 
Doug Johnson
Systems Developer/Engineer
The Ohio Supercomputer Center
djohnson at osc.edu