Re: [PATCH] sched: Enable arch-specific asym packing option in sched domain

[Michael Neuling <mikey@neuling.org>]

> Thanks for clarifying the Power7. I see the point - package vs cpu's within a package.
> 
> Yes, this is for Nvidia's ARM Quad-core Tegra CPU. It is a single
> package, organized flatly.  To save power, we want cores to be loaded
> up in order from cpu0, 1, etc.  The ASYM_PACKING flag seems to do
> exactly what we need if set in the domain flag.

Arrh, so it's not a performance issue but power savings.. that would
this asymmetric packing a lot more appropriate to use.

Mikey

> 
> Thanks,
> --Diwakar.
> 
> -----Original Message-----
> From: Peter Zijlstra [mailto:a.p.zijlstra@chello.nl] 
> Sent: Friday, May 04, 2012 3:30 PM
> To: Diwakar Tundlam
> Cc: 'Ingo Molnar'; 'Andrew Morton'; 'Christoph Lameter'; 'Michael Neuling'; 'Stephen Rothwell'; 'Benjamin Herrenschmidt'; 'David Rientjes'; 'linux-kernel@vger.kernel.org'; Peter De Schrijver
> Subject: RE: [PATCH] sched: Enable arch-specific asym packing option in sched domain
> 
> On Fri, 2012-05-04 at 15:18 -0700, Diwakar Tundlam wrote:
> > 
> > arch_sd_sibling_asym_packing() is already present under ifdef CONFIG_SMT.
> > I didn't touch that. I only added it to SD_CPU_INIT for all cpu's.
> > I assumed Power7 shouldn't use SD_CPU_INIT.
> 
> All archs use SD_CPU_INIT, its the default topology level for a package/socket. So now you've made Power7 prefer lower numbered sockets over higher numbered sockets.. not fatal, but not really nice either.
> 
> [ power7 is the only one that implements arch_sd_sibling_asym_packing ]
> 
> > Maybe I should define a separate weak symbol, say arch_sd_bias_to_lower_num_cpu()?
> > Then Power7 can define arch_sd_sibling_asym_packing() to be '1' and it 
> > will not break all-cpu init.
> 
> A slightly saner name would be:
> 
>   arch_sd_package_asym_packing()
> 
> FWIW, I suspect you're wanting to use this for some ARM chip (nvidia doesn't do much else -- aside from this graphics stuff) so that if there's hardly anything it runs on cpu0. Now, last time I checked, these ARM things had only 1 package, so I still don't see the point :-)
> 
> I suspect you want to modify SD_MC_INIT() with something like:
> 
>   arch_sd_mc_asym_packing()
> 
> Or is this the big-little thing and you're representing them as separate packages?
> 
> See how a little extra information avoids me having to endlessly second guess wtf you're actually wanting to do?
>