I am thinking of building a Beowulf of 17 dual-core nodes, one head node
and 16 computation nodes for material research simulation. I am also a
it's worth asking whether you need an interconnect more than gigabit.
my understanding is that many material sims can be statically partitioned
so that there is only modest communication between nodes assigned to
neighboring partitions.
1. Is Myrinet expandable?
I mean, if at first, I buy Myrinet for 16 nodes, should I buy more if I add
more nodes to the Beowulf?
for more ports, you can add line-cards later if your enclosure has space left.
since prices depend on the number of linecards, they jump % 8. so if you
want the head node (generally a good idea, I think) and 16 compute nodes,
then you'll need 3x 8pt cards, and the 3U chassis. adding nodes up to a
total of 24pt involves no change to the switch. going to the full 32pt
would mean another line card.
in my experience, the Myricom people are happy to help with configuration
questions.
2. When you do Monte-carlo or simulation, do you use commercial software like
MCNPX or Geant4? Which MD simulation and first principle package
would you recommend? Will it be good to install them all in one Beowulf?
my experience with MC is that they're very low on communication. the first
one we used was MCNP* (as it happens, on a myri 2g cluster).
regards, mark hahn.
_______________________________________________
Beowulf mailing list, [email protected]
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
