x86-64 on i386 (was Re: Promoting i386 version over x86_64?)
paul at dishone.st
Sun Dec 13 22:24:29 UTC 2009
On Sun, 13 Dec 2009, Chris Adams wrote:
> As soon as you bring in even one 64 bit user-space program that is run
> much, you've pulled in at least glibc and friends. At that point, you
> might as well run all (or as close to all as possible) 64 bit
> user-space, because the libraries are shared (code will be in the cache,
That's assuming that the footprint of libraries relative to distinct
applications is large enough to cancel out the space savings. (I have
no data either way). A 64bit kernel doesn't need any 32bit userspace.
An X server, on my 32bit system has about 8.5MB of programme text
(server and libs) and loads about another 1.5MB worth of modules
itself, i.e. 10MB.
So if you ran a 32bit system with a 64bit kernel and X server, you'd
lose out on about 10MB of shareable code. For comparison, my 32bit
system has O(10) times that allocated to things like browsers and
feed-readers. It's using 4.8GB in total (ex buffers/cache)
Space for text (programmes, code) is simply insignificant these days,
compared to the huge amounts of data which programmes allocate - data
which sometimes includes a lot of pointers.
You're also assuming that this cancels out the other benefits.
> The only time my systems have run 32 bit code in several years is for
> the Flash plugin (since the open-source plugins don't seem to be able to
> keep up and since the 64 bit Adobe plugin doesn't seem to get the
> security updates) and sometimes the Acrobat Reader plugin (since I've
> run into websites that assume they can embed PDFs in the page and AFAIK
> there's no plugin for Evince).
It's interesting that both you and drago have "almost always" (to
paraphrase) run 64bit pure systems. Surely that *reinforces* my point
about the futility of "64bit pure systems" as an achievable goal (in
the aggregate across all reasonable uses of a distro), and i386 being
a de-facto standard for software interfaces.
> As for the RAM overhead of 64 bit code vs. 32 bit code, I don't see it
> much in the real world. I have one 32 bit desktop at work, and
> comparing the resident RAM usage between it and a 64 bit desktop, I
> don't see much difference in the common desktop programs.
That's the wrong comparison - compare the aggregate RAM usage, with
each system in similar states.
> I know that for some reason PHP on 64 bit arches bloats up
> significantly (at least older versions), but that's the only major
> difference I've seen.
Pointer rich data structures, likely..
Anyway, as I don't intend to contribute anything, I'll try stop
Aside to the list: Thanks for all the hard-work on Fedora ;)
Paul Jakma paul at jakma.org Key ID: 64A2FF6A
Dogs just don't seem to be able to tell the difference between important people
and the rest of us.
More information about the fedora-devel-list