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

Re: Which kernels to include on ISOs? (Was: Re: Netboot Xen images for amd64)



On Friday 22 May 2009, Ian Campbell wrote:
> On Sat, 2009-04-18 at 12:49 +0200, Frans Pop wrote:
> > Also note that adding the 686-bigmem kernel to CD images still has a
> > rather high impact (as you'd need to add both the kernel udeb _and_
> > the regular kernel-image deb), which is rather undesirable for both
> > the netinst CD and the regular CD1. In the first case because of the
> > size increase it would cause [1], in the second case because it would
> > push packages important for the desktop task off CD1.
> >
> > Even the DVD1 image is getting tight as we currently support
> > installation of _all_ desktop environments from it. The margin for
> > businesscard images depends mainly on the capacity of actual
> > businesscard sized media.
>
> I've been thinking about this some more and I wonder if 486 + 686 is
> still the best option for DVD1 -- as opposed to 486 + 686-bigmem.

Yes, I'm very sure it is.

> IMHO the set of machines which benefit from a 686 kernel but are unable
> to run a 686-bigmem kernel is already small and getting smaller. I
> reckon those machine would be fine with a 486 kernel anyway, the 686
> optimisations don't buy you that much and SMP-but-non-PAE machine are
> an even smaller set (if such a thing even exists, I'm not sure).

I'm afraid I quite strongly disagree with you on this.
1) 686-bigmem has a significant performance penalty for systems that don't
   need it (otherwise it would be a non-issue as the option would just be
   enabled in the generic -686 kernel).
2) I guestimate that >90% of systems that really need a -686 kernel have
   less memory than the limit supported by the generic -686 kernel.
   Random sample: I have three desktops and a laptop running 686, and none
   of them comes even close to that limit.
3) IIRC there are fairly significant differences between -486 and -686
   performance on "normal" Pentiums and AMD boxes.
4) According to Linus, anybody who fits huge amounts of memory in "normal"
   Pentium systems is insane as the lowmem/highmem distinction will
   always continue to hurt you.
5) I would think that the "set of machines" you're aiming at is mostly
   64-bit capable (the 32-bit segment can hardly be said to be growing).
   In that case they really should be running either the amd64 arch, or,
   if they really want a 32-bit userland, i386, but with the -amd64
   kernel [0]! I expect that last would support Xen as well.

IMO your argumentation is strongly colored by your own goals here.

> New machines these days already have 1-2G as a pretty basic minimum and

And are 64-bit capable and thus shouldn't be using -686 at all!

> I predict that when squeeze arrives getting on for 4G will be a common
> default. A PAE kernel starts to become necessary around 3.5G anyway due
> to the PCI hole and even for machines with <3.5G of RAM you get things
> like NX support thrown in.
>
> The 686 kernel is still just an aptitude run away.

*shrug* so is the -bigmem kernel [1]. A more important issue for me is 
that we should make sure to install the correct *generic* kernel for 
regular users and leave specialized kernels to experts.


And now something slightly more constructive.

Personally I would say that full CD and DVD are not even very interesting 
for Xen installs: their content is desktop oriented, so why download a 
lot of shite you're not going to be using anyway?
Netinst and businesscard are much more relevant, but as explained earlier 
those have space restrictions.

But there is one image that might exactly fit the bill: the i386/amd64/ppc 
multi-arch netinst CD. Current size (Lenny): 488MB. For that it does not 
matter if it would grow a bit, and it's targeted exactly at your users: 
(semi-)professional sysadmins.

Only problem is that implementing adding Xen to just that image will 
require a fair few changes in debian-cd. In configuration, but I think 
also in code (you'll need to introduce a concept of variants within 
arches for D-I tasks). It'll not be trivial to implement that cleanly, 
though it should certainly be possible.
I'm afraid that I have no real affinity with Xen (the current discussion 
on lkml rather amuses me TBH), so I don't think I'd work on that [2].

Cheers,
FJP

[0] Supporting automatic selection of the -amd64 flavor in D-I for i386
    has never really been discussed. It would be interesting, but again
    the space limitations would have to be carefully considered.
[1] Yes, I know, that doesn't work if you want to install Xen.
[2] Though I would consider doing it for a suitable bounty.


Reply to: