> This reply is STRICTLY opinion on my part (please, no processor wars).
> In my opinion, if your application has a lot of parallelism (as I suspect
> given you are proposing a 100 node machine) then you are better off using
> Intel processors, strictly because they have the most market support, and
> thus are cheaper, easier to get, have more software available, and will be
> easier to upgrade in the future (unless Intel decides to keep changing the
> package every year <sigh>).
On the other hand, Alphas are much faster on FP, so he may find that
he gets much more bang for his buck with Alphas. "most market support"
doesn't mean you should check to see if Alphas are cheapest for his
application. Alphas are also easy to get, have plenty of free
software, and I'm not sure what upgradability means these days. Alpha
motherboards are the same size as everyone else's and they take
commodity memory. Intel is always changing their packaging, so I'd say
an Alpha's upgradability is about the same.
In short, worth a look.
> If you are already planning to use Myrinet, then the communication bandwidth
> isn't really an issue related to the CPUs.
Wrong. Some applications have higher bandwidth needs than others, so
this generalizatin is a bad one.
> Thus I am
> suggesting that 200 P2's are probably a better bet than 100 Alphas at the
> same cost (just an example, I'm not suggesting this is the actual cost ratio).
I think that if you do the math, you will find that this isn't the
case.
> Being a "dyed-in-the-wool" Beowulf person, I tend to believe that the best
> approach is to use a larger number of the most cost effective devices you
> can get. Cost effective including all aspects of hardware and software, not
> just CPU price and speed. For better or for worse, this is currently Intel.
I agree with you until the last sentence. If you look at the cluster
market, you'll find a lot of Alpha clusters.
-- g
