||Ten Tips for Building Your First High-Performance Cluster|
|Subject:||Use Identical hardware/avoid diskless|
I have build experimental clusters/multiprocessing setups using varied hardware (from diskless 386's to top of the range AMD)
I used it to play with ray tracing and animation so the data transfer rate was a lot less important than processing but you should try and make any problem map into this setup if you can. The elements in the cluster requested work from the root machine rather - if a machine went down no problem.
But while identical hardware may seem to be a management ideal I found that it was easy to manage a disparate collection of devices - if a 386 gave up then you just ditched it and the reliability of later machines is quite good.
If you isolate your cluster then the requirement to upgrade software on various machines is effectively limited to the software you want to run on your cluster. For a ray tracing or similarly easily arrangeable problem you'll find it a lot cheaper to just use whats around. If you want a massive distributed files system/database there may some advantages to be gained from lots of identical hardware but we need to look at the use of disparate hardware for clusters as it must be the cheapest way in the long run.
Hear us Roar