On 09/26/2013 09:28 PM, Arjan van de Ven wrote:

On 9/26/2013 6:42 AM, Srivatsa S. Bhat wrote:

quoted

On 09/26/2013 08:29 AM, Andrew Morton wrote:

quoted

On Thu, 26 Sep 2013 03:50:16 +0200 Andi Kleen [off-list ref]
wrote:

quoted

On Wed, Sep 25, 2013 at 06:21:29PM -0700, Andrew Morton wrote:

quoted

On Wed, 25 Sep 2013 18:15:21 -0700 Arjan van de Ven
[off-list ref] wrote:

quoted

On 9/25/2013 4:47 PM, Andi Kleen wrote:

quoted

quoted

Also, the changelogs don't appear to discuss one obvious
downside: the
latency incurred in bringing a bank out of one of the low-power
states
and back into full operation.  Please do discuss and quantify
that to
the best of your knowledge.

On Sandy Bridge the memry wakeup overhead is really small. It's
on by default
in most setups today.

btw note that those kind of memory power savings are
content-preserving,
so likely a whole chunk of these patches is not actually needed on
SNB
(or anything else Intel sells or sold)

(head spinning a bit).  Could you please expand on this rather a lot?

As far as I understand there is a range of aggressiveness. You could
just group memory a bit better (assuming you can sufficiently predict
the future or have some interface to let someone tell you about it).

Or you can actually move memory around later to get as low footprint
as possible.

This patchkit seems to do both, with the later parts being on the
aggressive side (move things around)

If you had non content preserving memory saving you would
need to be aggressive as you couldn't afford any mistakes.

If you had very slow wakeup you also couldn't afford mistakes,
as those could cost a lot of time.

On SandyBridge is not slow and it's preserving, so some mistakes are
ok.

But being aggressive (so move things around) may still help you saving
more power -- i guess only benchmarks can tell. It's a trade off
between
potential gain and potential worse case performance regression.
It may also depend on the workload.

At least right now the numbers seem to be positive.

OK.  But why are "a whole chunk of these patches not actually needed
on SNB
(or anything else Intel sells or sold)"?  What's the difference between
Intel products and whatever-it-is-this-patchset-was-designed-for?

Arjan, are you referring to the fact that Intel/SNB systems can exploit
memory self-refresh only when the entire system goes idle? Is that why
this
patchset won't turn out to be that useful on those platforms?

no we can use other things (CKE and co) all the time.

Ah, ok..

just that we found that statistical grouping gave 95%+ of the benefit,
without the cost of being aggressive on going to a 100.00% grouping

And how do you do that statistical grouping? Don't you need patches similar
to those in this patchset? Or are you saying that the existing vanilla
kernel itself does statistical grouping somehow?

Also, I didn't fully understand how NUMA policy will help in this case..
If you want to group memory allocations/references into fewer memory regions
_within_ a node, will NUMA policy really help? For example, in this patchset,
everything (all the allocation/reference shaping) is done _within_ the
NUMA boundary, assuming that the memory regions are subsets of a NUMA node.

Regards,
Srivatsa S. Bhat

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>