RFC: Page size on PPC/PPC64 builders

[Josh Boyer]

On 04 Mar 2008 16:37:29 -0600
Jason L Tibbitts III <tibbs at math.uh.edu> wrote:

> >>>>> "BP" == Brian Pepple <bpepple at fedoraproject.org> writes:
> 
> BP> It's already on the schedule for Thurs, since David asked me to
> BP> add it earlier today.
> 
> Could someone provide some information about why RHEL5 and FC6 use
> 64KB pages but later versions switched back to 4KB?

It essentially boils down to two things:

1) Comparative size of the machines
2) Workload

RHEL tends to be run on Enterprise class machines (imagine that), where
there is a very large amount of DRAM.  Using 64KiB pages lessens the
amount of kernel overhead needed to keep track of all those pages.  It
also improves MMU performance on those machines by causing fewer TLB
misses for the majority of workloads.  So for large Enterprise machines
and workloads, using 64KiB pages is a net win overall usually.

For Fedora, where a typical ppc64 machine is an Apple G5 with a
comparatively small amount of DRAM, 64KiB pages can actually cause more
issues than their worth.  You tend to run out of pages more quickly on
low memory machines since the smallest page size is 64KiB vs. 4KiB.
This can cause OOM conditions, etc.  There's also PS3, which only has
256MiB of DRAM, which means that you only have 4096 possible pages if
you use a 64KiB page size.  Not so good.  I get OOMs on my PS3 using
4KiB pages already.

The fact that FC6 was at one point switched to 64KiB pages was a
mistake.

Hope that clears things up for you (and FESCo) as far as the technical
issues behind 64KiB pages, etc.  Good question.  If there are more,
feel free to ask.

josh