Thread: Ubuntu i686 idea

if you are going to create such a group, count me in.

I have recently installed Gentoo on my machine, guess wat? It is sleek.
Speed, responsiveness, smoothness and much more.
However I'm not comfortable with Gentoo's package management system, which is why I came back to Ubuntu. I'm currently dual-booting.

We really need to tweak up ubuntu packages, may it won't be as good as Gentoo's performance, but something close to it is good enough.

[QUOTE=matt_hargett]First, newer instructions can help reduce the size of the binary if -Os is used instead of -O2.]/quote]Which isn't a performance benefit always. It normally is on GCC, however. But size vs. speed tradeoffs are usually just that on virtually all other compliers.

Extra registers enabled via -mmmx and -msse can also help increase performance by avoiding pointer dereference operations.

This doesn't help integer code one bit, except in some situations where they can be vectorized to SSE operations. MMX isn't interesting anymore (AMD64 doesn't even support it).

None of these would make support any more difficult than it already is since the existing binary packages mostly produce useless stack traces as-is.

Yes, they would. Mainly because GCC is frequently a less-than-perfect compiler.

If another repo were to be made, I would greatly prefer once that packaged more debuggable libraries.

They are packaged. Look at -dbg libraries.

I dont really think theres any good way to tell anything. You can benchmark all you want but still no matter what youll get slightly different results everytime. So many factors are involved with a computer. Heat being a major one try benchmarking when your system first boots up then try benchmarking after putting your cpu through some wear. You will notice a difference in the numbers.

And to sum up this thread in whole it was a waste 99% of it was childish bickering. While 686 packages would be cool it wouldnt be worth it for the ubuntu devs to do it. If you want to do it on your own system or even offer it to others go ahead thats the part that makes running linux great is the sharing.

Originally Posted by deathbyswiftwind

You can benchmark all you want but still no matter what youll get slightly different results everytime.

Not statistical meaningful differences.

Heat being a major one try benchmarking when your system first boots up then try benchmarking after putting your cpu through some wear. You will notice a difference in the numbers.

If you see a difference in the numbers due to heat it means you're engaging thermal throttling meaning that your data is instantly invalid. So no, heat doesn't matter.

Originally Posted by LordHunter317

MMX isn't interesting anymore (AMD64 doesn't even support it).

This isn't true (as far as I can tell anyway).

Code:

model name      : AMD Athlon(tm) 64 Processor 3400+
flags           : ... mmx fxsr sse sse2 ...

Either way, I don't understand why binaries are built for i386 anyway. Does anyone actually run a 386 anymore? Seems like you'd be hard pressed to even find someone using a classic pentium. Either way, the majority of PC users are gonig to be running on i686 machines, so why not have packages compiled for i686 by default and offer seperate i386 packages?

Ok, I've been watching this for a while and didn't see anybody posting *any* benchmarks, for crying out loud! So, I took the respectable dcraw.c (raw-to-png converter) and compiled it several times with different flags into separate executables. Then went to single user mode and killed all procs that resisted init. Then ran I 13MB-12MPixels raw image file from my S7000 3 times through each executable. I created a pure i386 binary, a i386 binary with sse2 support, a i686 binary, a i386 binary with pentium 4 tuning and a pentium4 binary with sse2 support. I sent all conversion output to /dev/null to avoid disk interference and I previously cp'ed all binaries and the raw image itself to /dev/null to have'em cached. The test machine was a P4/3GHz w/ 1GB of RAM (running Gentoo). Here are the results, the deviation of times in each 3 runs were so insignificant (less than 0.01 sec) that I post only one of each. [comments follow ]

# gcc -o dcraw.p4_sse2 -O3 -march=pentium4 -fomit-frame-pointer -fno-strength-reduce -msse2 -mfpmath=sse -lm -ljpeg dcraw.c
# gcc -o dcraw.i386_sse2 -O3 -march=i386 -fomit-frame-pointer -fno-strength-reduce -msse2 -mfpmath=sse -lm -ljpeg dcraw.c
# gcc -o dcraw.i386 -O3 -march=i386 -fomit-frame-pointer -fno-strength-reduce -lm -ljpeg dcraw.c
# gcc -o dcraw.i686 -O3 -march=i686 -fomit-frame-pointer -fno-strength-reduce -lm -ljpeg dcraw.c
# gcc -o dcraw.ubuntu -O3 -march=i386 -mtune=pentium4 -fomit-frame-pointer -fno-strength-reduce -lm -ljpeg dcraw.c

# time -p ./dcraw.i386 -c dscf0862.raf >/dev/null
real 10.26
user 13.34
sys 0.28
# time -p ./dcraw.ubuntu -c dscf0862.raf >/dev/null
real 10.26
user 13.34
sys 0.28
# time -p ./dcraw.i686 -c dscf0862.raf >/dev/null
real 9.57
user 12.42
sys 0.29
# time -p ./dcraw.p4_sse2 -c dscf0862.raf >/dev/null
real 7.09
user 9.15
sys 0.27
# time -p ./dcraw.i386_sse2 -c dscf0862.raf >/dev/null
real 7.28
user 9.39
sys 0.28

Comments:
a) The huge boost comes from the sse extensions (more than 30% over pure i386 in both P4 and 386 builds)
b) The i686 binary _is_ almost 10% faster than the i386 one
c) The "i386 binary tuned to P4 arch" is a myth of ubuntuland
d) 10% faster is well perceivable in heavy GUIs and apps
e) In slower machines the speed difference is even more obvious, that's why in my desktop P3/850, XGL runs smoothly in SSE optimized Kororaa and _very_ noticeably slower in Ubuntu.

My .02 euros.

Here's a relevant thread :
gen-bun-too (ubuntu + gentoo = genbuntoo)
http://www.ubuntuforums.org/showthread.php?t=208595

Originally Posted by nous

# gcc -o dcraw.p4_sse2 -O3 -march=pentium4 -fomit-frame-pointer -fno-strength-reduce -msse2 -mfpmath=sse -lm -ljpeg dcraw.c
My .02 euros.

what happens if you use -O2 or -O1 optimization settings?
I imagine Ubuntu is compiled using -O2 because it is potentially more stable.
Does that end up making much difference?

I did some looking into SSE, and it looks like you need a pIII or newer to use it. It's kind of a shame, the old pc's that could really use optimized execution dont have the ability to utilize many of the potential optimizations.

The only reason I even started looking into this was for an older PC I have that I want to give a boost to. What's the point now?