ARM hugetlb support?

ARM hugetlb support?

[Grazvydas Ignotas]

Hello,

We are developing an application that's drawing various geometry on
1024x1024 framebuffer (lines, triangles, etc) and is running on
Cortex-A8 CPU. It seems this kind of application is suffering badly
from TLB misses, as A8 only has 32 entry TLB cache.

I've found that hugetlb support patches have been posted some months
ago [1]. I've been testing them with very good results [2]. Are there
still plans to further develop them and eventually merge them?


[1] http://marc.info/?l=linux-arm-kernel&m=132912620601308&w=2
[2] Cycle measurements of rendering full scene (using cycle counter in
A8), x1000 cycles:

       normal hugetlb improvement
scene1 28507  21351   34%
scene2 11710  10138   16%
scene3 17178  13631   26%
scene4 4372   3810    15%
scene5 3525   3255    8%

-- 
Gra?vydas

ARM hugetlb support?

[bill4carson]

On 2012?08?20? 07:18, Grazvydas Ignotas wrote:
> Hello,
>
> We are developing an application that's drawing various geometry on
> 1024x1024 framebuffer (lines, triangles, etc) and is running on
> Cortex-A8 CPU. It seems this kind of application is suffering badly
> from TLB misses, as A8 only has 32 entry TLB cache.
>
> I've found that hugetlb support patches have been posted some months
> ago [1]. I've been testing them with very good results [2]. Are there
> still plans to further develop them and eventually merge them?
>
Hi, Grazvydas

I'm glad you got performance boost with my hugetlb patch!

It's a pity this patch series probably will not be developed.
I really don't have much time to continue with the development,
especially provide statistics how arm huge tlb could be tunned
well to produce better performance, that's based on much testing
and also application specific.

Anyway, query and discussion are always welcome.

>
> [1] http://marc.info/?l=linux-arm-kernel&m=132912620601308&w=2
> [2] Cycle measurements of rendering full scene (using cycle counter in
> A8), x1000 cycles:
>
>         normal hugetlb improvement
> scene1 28507  21351   34%
> scene2 11710  10138   16%
> scene3 17178  13631   26%
> scene4 4372   3810    15%
> scene5 3525   3255    8%
>

-- 
Love each day!

--bill

ARM hugetlb support?

[Steve Capper]

n Mon, Aug 20, 2012 at 00:18:12AM +0100, Grazvydas Ignotas wrote:
> Hello,
Hello Grazvydas,

> I've found that hugetlb support patches have been posted some months
> ago [1]. I've been testing them with very good results [2]. Are there
> still plans to further develop them and eventually merge them?
>

We are currently working on HugeTLB and Transparent HugePage support for both
LPAE and non-LPAE memory descriptors. The non-LPAE HugeTLB patches are based
on Bill's.

I hope to get some patches out to the list soon.

-- 
Steve

ARM hugetlb support?

[Grazvydas Ignotas]

On Mon, Aug 20, 2012 at 11:25 AM, Steve Capper <steve.capper@arm.com> wrote:
> n Mon, Aug 20, 2012 at 00:18:12AM +0100, Grazvydas Ignotas wrote:
>> Hello,
> Hello Grazvydas,
>
>> I've found that hugetlb support patches have been posted some months
>> ago [1]. I've been testing them with very good results [2]. Are there
>> still plans to further develop them and eventually merge them?
>>
>
> We are currently working on HugeTLB and Transparent HugePage support for both
> LPAE and non-LPAE memory descriptors. The non-LPAE HugeTLB patches are based
> on Bill's.
>
> I hope to get some patches out to the list soon.

Sounds promising. If you CC me when sending those patches, I can test
them on CortexA8 with my application.
BTW as you probably know know, Catalin seems to have hugetlb branch for LPAE:
http://git.kernel.org/?p=linux/kernel/git/cmarinas/linux-arm-arch.git;a=shortlog;h=refs/heads/hugetlb


-- 
Gra?vydas

ARM hugetlb support?

[Steve Capper]

On Tue, Aug 21, 2012 at 10:57:40AM +0100, Grazvydas Ignotas wrote:
> 
> Sounds promising. If you CC me when sending those patches, I can test
> them on CortexA8 with my application.

Will do; I too would be very interested in the performance difference.

> BTW as you probably know know, Catalin seems to have hugetlb branch for LPAE:
> http://git.kernel.org/?p=linux/kernel/git/cmarinas/linux-arm-arch.git;a=shortlog;h=refs/heads/hugetlb
> 

Yes thanks, I know.

-- 
Steve