Re: Bug#198158: architecture i386 isn't i386 anymore
On Wed, Jun 25, 2003 at 01:26:38PM +0900, GOTO Masanori wrote:
> Performance improvement is _not_ my primary intention.At least it needs to
> support libc6-686:
>
> - LinuxThreads floating stack support. It's ready for i686 and later.
>
> - NPTL/TLS support. NPTL currently supports i486 and later because
> pthread_spin_trylock uses cmpxchgl instruction (well, it's not
> difficult to support i386, but imagine pthread on i386 with the
> max clock (I recall it was 20MHz?) speed and memory...)
>
> and so on.
Ah, I was not aware of this. I am very interested in both of these
features.
> BTW, I also think that 5% performance improvement with optimization is
> valuable. It's not easy to accelerate 5% performance without any source
> code modification and hardware improvement. In addition, it reduces the
> user responce time, I think it's more important than increasing throughput
> (= computational speed).
5% may or may not be valuable depending on the cost. Where did this
particular figure come from (5%)?
> IMHO, the problem is the lack of real data which compares non- optimized
> vs optimized binary performance on the actual environment. I often heard
> the perfomance improvement issue using gcc optimization like Gentoo, but I
> merely saw the real perfomance comparison graph. So, supporting optimized
> library is not primary issue for at least libc.
I see a lot of handwaving from gentoo users and the like, but no convincing
concrete measurements. Most of what I see is:
- "I switched from <distribution X> to <optimized toy> and now <foo> is much
faster!"
- "I run <distribution X> on one PC, and <optimized toy> on the other, and
the optimized one is way faster! They're exactly the same hardware except
[they're not]"
- "I compared an un-optimized build of <foo> (no -O at all) to one compiled
with -O27 -pipe -fomit-frame-pointer -funroll-all-loops -march=pentium3
-mcpu=pentium3 -mfpmath=sse -mmmx -msse
-mother-options-i-dont-even-understand -mi-didnt-read-the-manual"
These claims are usually subjective, but sometimes they come with numbers,
which are meaningless because there are so many variables involved.
--
- mdz
Reply to: