Dear Tench:
Yes IBM wants to sell more high margin product. First, at $32 a transaction, they would not get any business at all. They probably meant $32 a transaction per minute. Multiplying that with the 440K TPM-C gets a total cost of $14 million. For a lousy 32 x 4 cluster. That is $450K a 4 way server. I know most of the cost is in the clients. Add 2 NICs to each client, use 24 port 100BT switches, add Beowulf Clustering software, public domain RDBMS, eliminate all the servers, and increase the clients (if you desire). Thus, you will get higher performance for a lot less money. If you lose a client, so what! The rest will keep going, and going, etc. And IBM is out a lot of money. The public does not participate in these tests (they cost too much up front money). But, more and more of them are switching to RAIC (Redundant Arrays of Inexpensive Computers) systems.
Since the way that TPM-C is setup, it could use even slow clients (read Durons, Celerons, or K6s) and still beat any server group. By converting the servers into ethernet switches (for a lot less per station), no matter what the servers prices were (unless they are really cheap (yeah, right)), the client clustering would take the performance crown.
So these benchmarks are really worthless except that PR and marketing departments want them. Buyer, beware!
Pete |
