discontigmem vs sparsemem on ia64 platforms

discontigmem vs sparsemem on ia64 platforms

[Nick Piggin]

Hi guys,

I wonder why you have DISCONTIGMEM_DEFAULT selected for GENERIC, SN2, and ZX1?
You don't explicitly disable SPARSEMEM for those platforms I think, so is
there another reason than bugs for this?

My followup question... I wonder why memory hotplug is disabled for
discontigmem? It seems like there is code there to handle it (athough
I don't know the memory hotplug code well so I didn't look too closely).

Thanks,
Nick

RE: discontigmem vs sparsemem on ia64 platforms

[Luck, Tony]

> I wonder why you have DISCONTIGMEM_DEFAULT selected for GENERIC, SN2, and > ZX1?
> You don't explicitly disable SPARSEMEM for those platforms I think, so is
> there another reason than bugs for this?

On ia64 DISCONTIGMEM was implemented first.  SPARSEMEM came later, and
nobody has sent patches to switch the config defaults over to use it.
I'd love it if somebody came up with some definitive arguments for
why SPARSEMEM can replace DISCONTIGMEM so we could clear out some of
the clutter that has accumulated by having both implementations. Last
time I looked (a very long time ago) it looked like SPARSEMEM would
have difficulties handling the extreme levels of sparseness on the
SN2 platform (high order physical address bits are used to route
memory traffic to the right node ... and since SN2 supports a lot
of nodes, the bits in question are way up high).

> My followup question... I wonder why memory hotplug is disabled for
> discontigmem? It seems like there is code there to handle it (athough
> I don't know the memory hotplug code well so I didn't look too closely).

Not at all sure of the history for this one.

-Tony

Re: discontigmem vs sparsemem on ia64 platforms

[Nick Piggin]

On Wed, Sep 02, 2009 at 09:56:40AM -0700, Luck, Tony wrote:
> > I wonder why you have DISCONTIGMEM_DEFAULT selected for GENERIC, SN2, and > ZX1?
> > You don't explicitly disable SPARSEMEM for those platforms I think, so is
> > there another reason than bugs for this?
> 
> On ia64 DISCONTIGMEM was implemented first.  SPARSEMEM came later, and
> nobody has sent patches to switch the config defaults over to use it.
> I'd love it if somebody came up with some definitive arguments for
> why SPARSEMEM can replace DISCONTIGMEM so we could clear out some of
> the clutter that has accumulated by having both implementations. Last
> time I looked (a very long time ago) it looked like SPARSEMEM would
> have difficulties handling the extreme levels of sparseness on the
> SN2 platform (high order physical address bits are used to route
> memory traffic to the right node ... and since SN2 supports a lot
> of nodes, the bits in question are way up high).

Not too sure. SPARSEMEM certainly can handle sparseness within a
node better than DISCONTIGMEM I think (discontigmem is basically
like a flat mem per-node so it gets redundant mem_map over holes
within a node).

I figured it was probably something to do with sn2, but I wondered
about bugs in the other cases too.

> > My followup question... I wonder why memory hotplug is disabled for
> > discontigmem? It seems like there is code there to handle it (athough
> > I don't know the memory hotplug code well so I didn't look too closely).
> 
> Not at all sure of the history for this one.

OK, I thought I'd ask here because some of the ia64 people are
involved in developing mem hotplug.

Thanks,
Nick