William Campbell (wcampbell@realworldtech.com) on 5/23/06 wrote:
---------------------------
>David Kanter (dkanter@realworldtech.com) on 5/23/06 wrote:
>---------------------------
>David, thank you for the review. Excellent work--and time consumming no doubt. ;)
>
>A general question for some feedback if i may. You make reference in your article
>on several occasions about future performance gains using tuning and optimised software.
>
>In the real world, how likely is that to happen though. My experience is that
>once a production platform becomes stable, the people who look after infrastructure
>are _loathe_ to change to (for example) a new JVM etc. because the validation and
>regression testing costs are prohibitive. Perhaps I've only worked in overly conservative
>organisations? So while the theoretical numbers are impressive, whether Woodcrest
>will be as dominant may not be so clear cut?

Well, I think it's a pretty safe bet that when Intel releases Woodcrest they will have new version of their compilers, math kernel libraries, etc. Their key partners (Oracle, MS, BEA, IBM, etc.) will all probably put in optimizations where relevant.

I think it's safe to assume that any compiled benchmark will show an improvement with new uarch aware compilers. JVMs and DMBS are all tested extensively in benchmarks, so it is reasonable to expect that the vendor will put in a lot of effort to maximize their score.

DK