dual AMD clusters

Bill Broadley bill at math.ucdavis.edu
Tue Jul 10 21:36:26 PDT 2001

Martin Siegert said:
> Also, the criterion is not whether two CPUs run your code twice as fast
> as one, but whether two CPUs on a dual run your code faster than two singles.

In our case it's more like, will a cluster of duals at $x have more
throughput with M duals or N single.  In our particular configuration
it doesn't actually decrease the node count much by adding a second cpu.

The overhead for our nodes is pretty large, once you add:
1.	100 mbit port (for nfs)
2.	Myrinet card
3.	Port on Myrinet switch
4.	Space in the rack (cost of rack, and cost of space for rack)
5.	Extra cooling
6.  Physical support (44 50 lb 1U nodes exceeds our load limit)
7.  Serial console/managment
8.  Power management 

The cost of an extra cpu per node ($250 ish for a 1.2 Ghz athlon MP), and
more expensive motherboard (an extra $250) wasn't that much.

For our codes performance increased more with a second cpu then the
corresponding cost.  So we decided to go with duals.

Bill Broadley
Mathematics/Institute of Theoretical Dynamics
UC Davis

More information about the Beowulf mailing list