Re: [gentoo-sparc] 64-bit gentoo-sparc, how to make it happen?

[Zhang Weiwu]

在 2006-06-25日的 19:50 +0200，Etaoin Shrdlu写道：
> On Sunday 25 June 2006 19:02, 张韡武 wrote:
>
> > Thank you for the answer. Just another smaller question: from pruely a
> > user's view (who simply wish to use it, rather then from developer
> > view who might gain knowledge and fun doing it), is it worthy to try
> > it (64bit sparc)?
> >
> > The original idea is the system as a print server is too slow (capable
> > of doing fast dithering for the inkjet printer, but not fast enough to
> > do real-time even-tone dithering to server the printer, usually cause
> > printer to wait for server), a rough gain of 20% to 30% CPU power, by
> > my guessing of observing top(1), should be enough for the printer.
> > Dithering is always a CPU intensive job and we do a lot of printing.
> >
> > I observed kernel is taking only a very small percentage of CPU usage,
> > if I try hours to get 64-bit kernel running but only obtain 5% speed
> > gain, that may not be worth. Or should I try Solaris 10? Should
> > Solaris provide better performance in this very case? (how much
> > performance gain might happen? 10%? 20%?) I don't need a very close
> > estimation, just if someone can give me a very rough estimation is
> > very helpful, for I am not techincally able to do the decision:)
> >
> > If switch to Solaris for performance (no offensive, I am a 5 year's
> > hardcore Linux user), generally speaking, is Solaris very compatible
> > with PCI cards like NEC USB controller, Promise IDE controller and
> > ethernet cards?
>
> Well, I'm not a sparc dev so probably I cannot give you the most accurate
> answer you could get; moreover, I cannot speak about solaris since I
> don't know it. Anyway...
> From what I've read, the main reason gentoo stays 64 bit kernel/32 bit
> userland is that switching to a full 64 bit system (perhaps with
> multilib) will not be a considerable improvement, but rather just a
> waste of memory. This is different from what happens in the x86 world,
> where being 64-bit means getting a lot of benefits (like, for example,
> many more cpu registers). In the sparc world, for an userland app "being
> 64-bit" means, basically, "wasting more memory" (due to larger
> executables binaries). In short, I think that (from the pure user's view
> you were talking about), it's not worth trying.
> However, I believe that a (highly experimental) userland 64-bit profile
> exists (but of course don't expect that to work flawlessly or to be
> usable in a production environment).
> Read this thread on the forums, which explains all:
>
> http://forums.gentoo.org/viewtopic-t-442194.html
>
> btw, AFAIK solaris uses a similar 64bit-kernel/32bit-userland approach
> (or, at least, *most* of the userland is 32bit).
>
> Hope this helps (and please somebody correct me if what I said is wrong).

Blame me for being stupid, but I simply cannot understand. Truly in x86
world 64-bit is often marketed as much faster then 32-bit applications,
if this is not ture in sparc system, the question is why people bother
to build 64-bit stuff, being slower and wastes more memory at all? For
sure I am not an arch guru, but there must be a good reason behind it.

I am one of stupid graduated CS student the modern universities
produced, I can understand perfectly well that 32-bit binary
applications perform equally on 32-bit or 64-bit platform, but I was
tought that is because the binary is 32-bit: if the source is compiled
targeted to be 64-bit binary, usually it is faster especially for CPU
intensive tasks, the compiler optimize the source and use 64-bit when
possible (perhaps especially for float point). In my case I can
re-compile gutenprint or cups to generate 64-bit binary. Especially the
job like dithering for the printer. In my understanding dithering is
purely computation of how to mix colors, few I/O and only some memory,
that looks really like something can be optimized by using 64-bit.

For the suggestions I can easily take all of them, just now I am
curious:) If 64-bit is truly considered an enhancement for CPU, there
must be some use of it, and if purely CPU intensive tasks (dithering)
cannot benifit from enhancement for CPU, whatelse can take the advantage
of 64-bit? (Okay, now I am not in purely user's view, I wish to dig into
it a little bit because that makes me really curious).

-- 
gentoo-sparc@g.o mailing list