Re: hardware

[Rob Latham <rob@terizla.org>]

Robert Tompkins (tompkinr@clarkson.edu) said:

> I am building a cluster and i was wondering what would be more important,
> memory with lower latency (cas 2 compared to cas 3) or more memory( 512
> megs compared to 1 gig). Is memory latency a bottleneck in clustered
> computering or is it usually memory size that is the bottleneck. 

depends on your problem

> Also would it be better to have 20 smp machines or 40 single
> processor machines( both of which would be connected by
> 100base-t).

depends on your problem

w/o knowing more about what you want to do with your beowulf, it's
very hard to answer the (very good) questions you ask.  It's why my
employer is in business.  

*generally* you want fast memory.  unless you are working with data
structures so big that you might touch swap.  

*generally* you will never see 100% utilization of the second cpu in
an smp node.  60% is more normal, but i've seen it get as high as 95%.
two cpus have to contend for the memory bus, and on x86, the memory
bandwidth is already pitifully small.  2 cpus have to contend for the
one network pipe going into the node.  basically every shared resource
represents a *potential* bottleneck ( only your applcation can say if
it really is a limit).   sometimes jobs running on the 2 cpus can
share memory ( some incarnations of mpich support this, i belive, if
you configure with the right flags ),  but not often.

If you can, get 4 nodes: permute smp and up with slow ram and big ram.
it won't reflect the communications overhead, but you can start to
measure the effect hardware has on your problem: you might find the
slow ram to have a negligible impact on run time for example.  

good luck
==rob

-- 
Rob Latham: linux A-Team                          Bethlehem, PA USA
EAE8 DE90 85BB 526F 3181                   1FCF 51C4 B6CB 08CC 0897