x86-64 on i386 (was Re: Promoting i386 version over x86_64?)

drago01 drago01 at gmail.com
Sun Dec 13 20:09:17 UTC 2009


On Sun, Dec 13, 2009 at 8:16 PM, Paul Jakma <paul at dishone.st> wrote:
> On Sun, 13 Dec 2009, drago01 wrote:
>
>> such a setup does not make much sense, when your hardware supports x86_64
>> not using it for userspace is a waste ....
>
> a) i386 has a lower memory footprint, as has been mentioned in this
>   thread.

Yes which is pretty much the only valid complaint but trading memory
for performance is a price I am willing to take ...

> b) The amount of code on your system that is CPU bound and/or
>   memory-bound due to register pressure, to an extent that the x64
>   registers would make an appreciable difference is probably not
>   that significant
>
>   - kernel hotpots

The kernel doesn't do any have computing...

>   - graphics hotspots (X server perhaps)
>
>   I havn't measured this, but nor have the people who say x86_64 is
>   faster AFAICT, and there's plenty of experience to say that most
>   software is far from CPU bound or memory bound.

Yes but the stuff adds up, you gain almost nothing by running i686
code but where it matters x86_64 can make a HUGE difference.

> c) There is a definite cost to a distro in having to maintain 2
>   x86_64 and i386 as separate arches

Not a reason to move forward with hardware development.

>
> d) Like or not, i386 is the de-facto standard for binary interfaces:
>
>   - Netscape plugins

This is slowly being fixed.

>   - Windows executables

Nobody stops you from running i386 apps on a x86_64 system.

>   - VM images to run in, say, QEMU/KVM
>   - Sandboxing technologies for, say, browser plugins (I think
>     Google have stuff in this area)
>   - Free software windows-only apps (don't know if they exist)
>
>   All the code here can be open-source/free-software and still be
>   relying on i386 as a widely known and hence convenient
>   /interface/. As such, it likely needs to be supported on x86_64
>   kernel-based systems anyway, as performantly as possible. (And
>   yeah, I gather KVM x86_64 doesn't work for i386 VMs - annoying).

Er.. don't quite get your point here, what is stopping me from running
i686 VMs on a x86_64 host?
I have been doing this for a while  and there are there problems (you
don't even need multilib for that)

> So personally I think x86_64-pure is unrealistic and, independently, I think
> 32-on-64 makes sense, but hey. :)

I did not suggest using pure x86_64 but using x86_64 where we can (ie.
not just the kernel).




More information about the fedora-devel-list mailing list