32 + 8 nodes beowulf: thank you!

Paul Maragakis plm at hdsc.deas.harvard.edu
Wed Feb 14 08:35:19 PST 2001

Hi everyone,

Thank you for the overwhelming response to the draft.  I read all of your
mail with great interest and I will commit many of the changes that you
proposed.  Sepcial thanks go to Jeff, Brian, Jared, Serguei, and Joseph.  
After I complete the rewrite and the target is clear I will contact some 
of the vendors that showed interest.  This will take another ten days
because of other urgent matters.

Brief list of changes:
*) The 2 hubs will go; we will get a second switch instead.  
*) Channel bonding will be used. 
*) A GBit ethernet for switch to master will be considered.  This will
need more homework to make sure things will work out as expected.  (will
channel bonding mess this thing up?)
*) A tape with autoloader will be selected.  All recommentations welcome.
*) A second SCSI card for the tapes will be used.
*) Cooling and noise insulation will be explicitly mentioned in the draft.
*) KT 133A, 133 MHz memory, and the 133 MHz FSB Athlon.
*) NIC types will be explicit, maybe the Intel EEPro100, or 3Com 3C905.
*) Network booting will be reconsidered.  The key idea was to simplify
kernel upgrades.  There are other ways of doing so and they may be better.
*) The chassis will be explicitly mentioned.
*) The key component prices (memory, CPU) will be checked by talking to
vendors directly.  

On the software side (which is not needed as part of the proposal but I
guess will take the most time to make right):
*) The PGI-CDK will say: we run a 16-node parrallel HPF code and only
this one compiles it.  It will be made clearer.  We only need 1 copy of
the software.
*) PVFS, LVM, NovaNet 8, MPI/Pro will be considered.

More information about the Beowulf mailing list