[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: Wishlist: Source-based option



On Wed, 2003-03-19 at 07:23, David Krider wrote:
> Well, frankly, if you were talking about me in particular being an
> "easily influenced geek," I do take some mild offense. What you said
> sounds really... professional, but in my case, you're wrong. I fought
> the idea for weeks. My buddy put Gentoo on every Linux box he had
> (something like 8 or 10), and kept extolling the virtues of the speed. I
> kept telling him he was full of crap. When I finally tried it, I didn't
> want it to be true, because I didn't want to spend all my time
> recompiling. The bottom line is that I'm saying it's 10 or 15 percent
> "snappier" based on a negative presumption.

"snappier" isn't a benchmark I can use in a quantiative analysis.
And even if you did get hard numbers on gentoo's performance for
specific tasks...you still have to be careful that you are measuring
compile optimization issues and not configuration issues.  You can't
just line up 2 distros side by side and say with great certainty that
the compiler optimizations are THE difference, and not silly things like
prelink being used.  It makes far more since to test that idea by
rerolling yer Red Hat srpms with just compiler option changes, and
benchmarking that difference, leaving everything else the same in terms
of configuration. If you really cared about making quantiative
benchmarks instead of this "feels faster" stuff.  And, I'd be very
surprised if Red Hat hadn't done a modicum of benchmarking...since they
do provide kernel and glibc arch specifc packages, and I doubt they just
decided on that without some hard numbers the backup the arguments.

> If people don't want to believe me, fine: don't believe me. 
NUMBERS!!!!! I will believe hard numbers, produced by a scientifically
repeatable method which does its best to keep all other factors, other
than the one being tested the same. Gentoo could be doing a vast array
of things that have nothing to do with actually recompiling the source
code. I'll believe you..as quickly as I believed the cold fusion
researchers...when either of you come up with a reproducible scientific
approach. 

> I'm telling
> everyone that there's noticeable performance gains to be had using it,
but you have no idea WHY.

> and I'm only saying that to encourage people to try it out. If you
> disagree with me, don't just say "balogna" and claim misguided
> psychological issues are involved. Try it for yourself. Put it on a
> machine you're in front of a lot, and just take a test drive. 

I have better things for my cpu to do for a week instead of compiling up
a system...like using it to email.

> I don't
> know of any way to officially benchmark this stuff, else I would, and
> prove all the naysayers wrong. ;-) 

Even if you could figure out how to run benchmarks..I still don't think
you'd have any idea how to figure out what exactly gentoo is doing
differently. Gentoo is not Red Hat with different compile options. I
have a sneaky suspicion that library prelinking is heavily involved in
that "snappiness" you talk about.
http://www.gentoo.org/doc/en/prelink-howto.xml
"Support is built into Portage." Prelink has very very little to do with
compiling from source. But prelink does barf on some binaries, like
those pesky nvidia drivers yer having problems with. If gentoo is trying
to prelink that nvidia drivers...well no wonder they don't work for you.

> Hey, you've got a machine that runs Phoebe, don't you? Why else are you
> on this list? Obstensibly, that's not a production machine. Try it
> there, and then you won't have to guess at what's possible.

I'd rather spend 3 days actually bughunting on the phoebe box while I do
day to day tasks, than spend 3 days compiling gentoo, just to end up
with binaries that use prelinking to give me a feeling of speed. Have
fun burning clock cycles on recompiling dependancy chains when major
updates come down the line...make sure you factor in the source code
compile time into your "performance" metric.

-jef"prelink is cute...but i think its one of those better know what the
hell you are doing kind of CONFIGURATION issues"spaleta

Attachment: signature.asc
Description: This is a digitally signed message part


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]