[Fedora-xen] Heads-up: Requiring PAE for running Xen

[Jeremy Katz]

On Wed, 2006-05-17 at 21:31 -0600, Lamont R. Peterson wrote:
> Personally, I like 4 better.

Well, yes.  But that requires someone to actually do the work.  And it's
*hard*.  And the pragmatic side of things means we need to have
something sooner rather than later

>   In addition, if someone wants to turn on PAE for 
> their 32bit Xen boxes, it's not that hard to rebuild your own kernel RPMs.  
> It only takes 6.5 hours on this old-school mobile P4 notebook, and on my Dual 
> 1.6GHz Athlon, it's about 1 hour to build a complete set of FC5 kernel 
> packages.

Yes, but then you have to rebuild install trees too.  Your hv + dom0
kernel + domU kernels all have to match.  And that's not really a
position that can be sanely supported.  If someone reports a bug, how do
you know if it's due to the fact that they recompiled with PAE[1] or not

> However, I have another question about PAE; let's say I have a box with 2-4GB 
> RAM and I run a non-PAE kernel (Xen or otherwise, but I am using Xen, of 
> course :) ) ... what kind of performance hit should I expect to take from 
> turning on PAE?  I've heard numbers around 3:1, before, but that doesn't 
> quite sound right to me.

The numbers shouldn't be anywhere near that bad.  Also, keep in mind
that until just recently, the smp kernel has required PAE for as long as
we've been doing 2.6 kernels.

> Now, let's say I scale up the RAM in that box to 8GB after turning on PAE.  
> Will there be a difference in memory access performance compared to using PAE 
> with 4GB or less?

This is where you start seeing impact.  What the impact is depends a lot
on your workload, though.[2]

> Lastly, are you talking about 16GB or 64GB support.  As I understand things, 
> there are almost no 32bit processors (if any) that can actually use PAE to 
> access RAM past 16GB.

Xen only currently supports up to 16 GB

Jeremy

[1] Which will tickle different bugs
[2] Based on my understanding from people who have run benchmarks.