I'm enjoying this dialog. I agree with Shmuel that if it were totally simple to estimate capacity requirements across model generations then each new model would be quite disappointing. And we wouldn't even be able to do certain things in real world volumes such as cryptography, which is rather important of course. Processor performance improvements often yield varying benefits depending on the particular workloads.
I suppose then there's the question of whether divergence from capacity and performance forecasts matter. Sometimes yes, sometimes no. We deal with statistical means, medians, standard deviations, distributions, and percentiles all the time in our everyday lives. For example, automobiles have published miles per gallon (or liters per 100 Kilometers) fuel efficiency ratings, typically one number for stop-and-go city driving and another for highway driving. Your results will probably be different than the published numbers, resulting in higher or lower fuel costs and consequent environmental impact. But those differences may or may not matter. (They matter less in the U.S. than in Belgium, ceteris paribus.) Manufacturers and regulators are now struggling with how and whether to publish such numbers for hybrid and electric vehicles. In other words, there's a significant technology breakthrough in vehicle design, and the existing efficiency metric is at least somewhat harder to apply and to interpret. (Obviously banning hybrid and electric vehicles because it's harder to assign them single efficiency numbers would be ridiculous.) I also agree that sometimes you need a little expert help in such matters. (Maybe even a trial or benchmark.) The amount of help should be in proportion to the risks and rewards, though. Vehicle fleet buyers probably need to worry more about fuel consumption results than lower volume buyers, for example. And sometimes that help isn't really technical. I can't even begin to list all the times when someone expends a great deal of effort trying to optimize some aspect of computing when the actual, real, best solution involved somebody just agreeing to change a number in a spreadsheet or contract. Or when somebody deploys armies (comparatively speaking) to tweak their mainframe to squeeze out 0.02% more throughput at the same time that they're spending (and wasting) millions on everything else flooding into their data centers. Maybe (maybe) they should do both, but there are priorities! Then again, I've occasionally been known to spend countless hours shopping for a US$20 product in search of a coupon, promotion, and/or rebate to drive the price all the way down to $19.68. I also couldn't resist buying two cans of pumpkin at 19 cents each simply because ... they were only 19 cents each! Those cans then sat on my shelf for two years until I gave them away. We humans aren't always rational, are we? :-) -------------------------------------------------------------------------------------------------------- Timothy Sipples Resident Enterprise Architect (Based in Singapore) E-Mail: [email protected] ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
