At 10:22 AM 4/24/00 -0700, [EMAIL PROTECTED] wrote:
>Dan Sugalski wrote:
>
> > Dunno how many folks have machines with the newer toys in them, but I just
> > ran perlbench against 5.6.0. It turns out that if you specify an
> > /ARCH=HOST, and compile on an EV6 system, perl gets an overall boost of
> > about 5% over a plain build. Some of the tests show a more significant
> > boost (the string/base64 test got a 40% boost over a plain compile, and
> > string/tr jumped 70%), generally those that deal with strings, which makes
> > sense given the extra instructions in the EV5+ processors.
> >
> > The sad thing is that, even with the extra goodies, 5.6.0 was only 15%
> > faster than 5.004_04, and that was with 5.6.0 racking up a numeric sort
> > result 4x that of 5.004_04. (Plus 5.6.0 was built with the Dec C 6.2, 
> while
> > 5.004_04 was built with, I think, Dec C 5.5...)
> >
> > I don't think we're quite pulling ahead here, but I don't know exactly
> > where 5.6.0's losing it's performance.
>
>Interesting result (did you post to p5p?).

While I've been posting to p5p, it's not been about this... But that's a 
separate issue.

>I do not quite understand
>the combination of the first sentence in the second to last paragraph
>(where 5.6.0 is 15% faster) and the last sentence where you seem to
>imply 5.6.0 is not quite pulling ahead.  Isn't a 15% speed improvement
>a speed improvement?

Sort of. It depends on the tests and how the final numbers are calculated. 
Perlbench has a whole slew of tests it runs and, in general, 5.6.0 is 
slower on the tests (70-80% of them) than 5.004_04 is. There are a few 
outliers that pull the final numbers in 5.6.0's favor, such as the string 
and sort tests. The big question is, of course, how representative are the 
numbers? Does a win in the character handling and sort stuff buy you a win 
overall, or do the slowdowns on everything else weigh in more heavily? It's 
hard to tell.

5.6.0's performance is what I'd call lackluster, honestly. (Though the 
Win32 folks have it even rougher, as the fork emulation requirements suck 
down 15% of its speed relative to a plain 5.6.0 build)

                                        Dan

--------------------------------------"it's like this"-------------------
Dan Sugalski                          even samurai
[EMAIL PROTECTED]                         have teddy bears and even
                                      teddy bears get drunk

Reply via email to