On Sun, Nov 09, 2003 at 03:37:11PM +0100, Eduard Bloch wrote:
| #include <hallo.h>
| * Michael Poole [Sun, Nov 09 2003, 09:22:13AM]:
| > Eduard Bloch writes:
| >
| > > Do you see now that 8 of your 10 percent come directly from the
| > > application code and other two maybe from the optimized libc? There is
| > > not{hing| much} we have won using an optimised kernel. But the placebo
| > > effect has been demonstraded once again.
| >
| > You have not shown what you claim you have shown. You have shown that
|
| No. Please read my initial mail then and what Glenn tried to proove with
| his bzip2 test.
I believe that Michael is correct. A summary of the messages leading up
to this one on the thread is as follows: [with my comments in square
brackets]
Nikita: There are significant performance gains optimising the
kernel for a modern CPU rather than i386.
Eduard: Optimising kernel code doesn't help as other hardware is the
limiting factor.
Nikita: No, you're wrong.
Andrew Suffield: Prove that optimising for a particular CPU makes things
faster rather than slower.
Glenn: It /does/ make things faster, really! Here are some pretty
numbers!
[Provides an example of Gentoo's bzip2 vs Debian's bzip2, running on an
identical kernel. Gentoo's is ~10% faster. Thus Glenn has given one
valid data point of code running faster when optimised for the machine
it is running on.]
Eduard: Glenn's comparison [of user-space code] is invalid as his
kernel is optimised in both cases. Provides an example of bzip2 running
on a P4 kernel vs a "vanilla" kernel, with no significant performance
difference. Therefore optimising the kernel yields no performance
advantage.
[Output of "time" shows that an insignificant proportion of bzip2's run
time is spent in kernel space anyway, thus any difference in kernel
optimisation is not going to make much difference to bzip2's run time
anyway. Correct conclusion is that kernel optimisations make little
difference to some CPU-intensive apps, not that kernel optimisations
make little difference to anything, ever.]
Glenn: I was comparing user-space optimisation so the kernel is
irrelevant.
[i.e. Glenn's point was about optimisation in general, rather than
optimisation of the kernel in particular. Which is okay 'cos he was
replying to Andrew Suffield, and that's what Andrew seemed to be talking
about too.]
Eduard: Yeah, user-space optimisation is the only thing that matters,
here's another "benchmark" [more or less the same as Glenn's]. See,
I was right!
[Eduard's comparison here provides another data point in favour of
optimisation making a difference in bzip2. Once again utterly
irrelevant for benchmarking different kernels. If one wanted to compare
the effects of optimising the kernel, a valid benchmark would be one
which actually spends most of its time executing kernel code rather than
user code.]
Michael and Cameron: No-one has shown anything remotely interesting
here about the effects of optimisating kernel code. Eduard's
interpretation of his own benchmark is invalid.
Eduard: No it's not!
Cameron: Yes it is!
(Repeat ad nauseum...)
| I know, and that was NOT my claim (but the counterpart!). Don't put
| words in my mouth.
Heh. Now you'll probably accuse me of doing the same. So: tell me
where I have misrepresented you in the above.
Cheers,
Cameron.
