[Beowulf] Which is better; 4 Mountains or 5 Hills
Robert G. Brown
rgb at phy.duke.edu
Sun Apr 1 21:43:13 PDT 2007
On Sat, 31 Mar 2007, amjad ali wrote:
> Hi All,
> Which one of the following two choices are better (assuming both
> clusters have nearly same cost):
> 1) A 4-node cluster having 8 AMD Opteron of 2.6GHz each and 2GB RAM/node
> 1) A 5-node cluster having 10 AMD Opteron of 2.2GHz each and 2GB RAM/node
What is the meaning of this word "better"? Better for what?
4x2.6 = 10.4 aggregate GHz (whatever that means)
Can do a distributed calculation that fits into 8 GB RAM
Fewer faster nodes means less Amdahlian penalty for parallel code.
Probably uses less power, has slightly lower aggregate maintenace
5x2.2 = 11 aggregate GHz
Can do a distributed calculation that fits into 10 GB RAM
More slower nodes means a bit more flexibility and less penalty if a
node is down
Probably uses more power, has higher costs.
Which is "better"? B) has more aggregate GHz (and presumably FLOPS by
at least some measures) so it is better. It can also hold a larger
partitioned problem, which for some people makes it MUCH better, and for
others just wastes money. A) might well run faster for some kinds of
parallel code, and it probably costs less in the long run, once
(expensive) human time and infrastructure costs are accounted for.
> Consider GiGE (max 1000mbps) as the interconnect.
> Is there any significant performance difference between the two?
Sure, maybe, for some code. For example, try running a 9 GB partitioned
computation on A and oooo, slow (as it swaps to run at all). Try
running a computation with lots of little messages, but there are just
plain fewer of them between four nodes than for five, and A might shine.
> Does the compasison depends on the MPI communication among nodes, for
> a given application?
It depends on EVERYTHING. The task mix, how the task(s) are executed,
whether, how, and how much they intercommunicate during execution, how
big they are (or might be) and whether four boxes plugged into your only
available circuit doesn't blow the breaker where five does.
> Thanking all in anticipation.
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb at phy.duke.edu
More information about the Beowulf