On 10/26/10 04:16, Hearns, John wrote:
I have worked as an engineer for two HPC companies - Clustervision and Streamline. My slogan phrase on this issue is "Any fool can go down PC World and buy a bunch of PCs"
Well if you are buying PCs in bulk at retail pricing, you are a fool anyway. Plus most PC World PCs won't have ECC RAM so I wasn't really referring to those as few of us tolerate random bit flips.
However, as regards price, I would say that actually you will be paying very, very little premium for getting a supported, tested and pre-assembled cluster from a vendor. Academic margins are razor thin - the companies are not growing fat over academic deals. They also can get special pricing from Intel/AMD if the project can be justified - probably ending up at a price per box near to what you pay at PC World.
Again, not comparing PC World to Tier 1 bulk purchases. I'm comparing Tier 1 bulk purchases w/o an OS (so you can DIY) with specialized HPC vendor purchases where you don't have to DIY. Even then, perhaps it breaks even the first year if you get a very, very good deal from the HPC vendor. However, to get the deal you are probably contracted into four or five years of support and when considering HPC, involving more humans are the fastest way to get a really inefficient and expensive cluster. After the first year and up until the lifetime of the cluster involving human support annually will add a large cost overhead you have to account for at the beginning (and probably buy less hardware because of which).
Or take (say) rack top switches. Do you want to have a situation where the company which supports your cluster has switches sitting on a shelf, so when a switch fails someone (me!) is sent out the next morning to deliver a new switch in a box, cable it in and get you running?
That's probably a hell of a lot faster than waiting on a vendor to get you a new switch through some RMA process. Plus you know the cabling is done right :).
Optimally IMHO, in university setups physical scientists create the need for HPC. These types shouldn't (as Kilian mentions) need to inherit all of the responsibilities and overheads of cluster management to use one (or pay cluster vendors annually for support). They should simply walk over to the CS department, find system guys (who would probably drool over the potential of administering a reasonably sized cluster) and work out an agreement where the physical science types can "just use it" and the systems/CS guys administer it and can once in a while trace workloads, test new load balancing mechanisms, try different kernel settings for performance, etc. This way the physical scientists get their work done on a well supported HPC system for no extra cash and computer scientists get great, non-toy traces and workloads to further their own research. Both parties win.
Now in organizations that don't have a CS department I agree that HPC vendors are the way to go.
ellis _______________________________________________ Beowulf mailing list, [email protected] sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
