Currently max_sane_readahead() returns zero on the cpu having no local memory node
which leads to readahead failure. Fix the readahead failure by returning
minimum of (requested pages, 512). Users running application on a memory-less cpu
which needs readahead such as streaming application see considerable boost in the
performance.
Result:
fadvise experiment with FADV_WILLNEED on a PPC machine having memoryless CPU
with 1GB testfile ( 12 iterations) yielded around 46.66% improvement.
fadvise experiment with FADV_WILLNEED on a x240 machine with 1GB testfile
32GB* 4G RAM numa machine ( 12 iterations) showed no impact on the normal
NUMA cases w/ patch.
Kernel Avg Stddev
base 7.4975 3.92%
patched 7.4174 3.26%
Suggested-by: Linus Torvalds <[email protected]>
[Andrew: making return value PAGE_SIZE independent]
Signed-off-by: Raghavendra K T <[email protected]>
---
I would like to thank Honza, David for their valuable suggestions and
patiently reviewing the patches.
Changes in V6:
- Just limit the readahead to 2MB on 4k pages system as suggested by Linus.
and make it independent of PAGE_SIZE.
Changes in V5:
- Drop the 4k limit for normal readahead. (Jan Kara)
Changes in V4:
- Check for total node memory to decide whether we don't
have local memory (jan Kara)
- Add 4k page limit on readahead for normal and remote readahead (Linus)
(Linus suggestion was 16MB limit).
Changes in V3:
- Drop iterating over numa nodes that calculates total free pages (Linus)
Agree that we do not have control on allocation for readahead on a
particular numa node and hence for remote readahead we can not further
sanitize based on potential free pages of that node. and also we do
not want to itererate through all nodes to find total free pages.
Suggestions and comments welcome
mm/readahead.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/mm/readahead.c b/mm/readahead.c
index 0de2360..1fa0d6f 100644
--- a/mm/readahead.c
+++ b/mm/readahead.c
@@ -233,14 +233,14 @@ int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
return 0;
}
+#define MAX_READAHEAD ((512*4096)/PAGE_CACHE_SIZE)
/*
* Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
* sensible upper limit.
*/
unsigned long max_sane_readahead(unsigned long nr)
{
- return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
- + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
+ return min(nr, MAX_READAHEAD);
}
/*
--
1.7.11.7
On Tue 18-02-14 12:55:38, Raghavendra K T wrote:
> Currently max_sane_readahead() returns zero on the cpu having no local memory node
> which leads to readahead failure. Fix the readahead failure by returning
> minimum of (requested pages, 512). Users running application on a memory-less cpu
> which needs readahead such as streaming application see considerable boost in the
> performance.
>
> Result:
> fadvise experiment with FADV_WILLNEED on a PPC machine having memoryless CPU
> with 1GB testfile ( 12 iterations) yielded around 46.66% improvement.
>
> fadvise experiment with FADV_WILLNEED on a x240 machine with 1GB testfile
> 32GB* 4G RAM numa machine ( 12 iterations) showed no impact on the normal
> NUMA cases w/ patch.
Can you try one more thing please? Compare startup time of some big
executable (Firefox or LibreOffice come to my mind) for the patched and
normal kernel on a machine which wasn't hit by this NUMA issue. And don't
forget to do "echo 3 >/proc/sys/vm/drop_caches" before each test to flush
the caches. If this doesn't show significant differences, I'm OK with the
patch.
Honza
> Kernel Avg Stddev
> base 7.4975 3.92%
> patched 7.4174 3.26%
>
> Suggested-by: Linus Torvalds <[email protected]>
> [Andrew: making return value PAGE_SIZE independent]
> Signed-off-by: Raghavendra K T <[email protected]>
> ---
> I would like to thank Honza, David for their valuable suggestions and
> patiently reviewing the patches.
>
> Changes in V6:
> - Just limit the readahead to 2MB on 4k pages system as suggested by Linus.
> and make it independent of PAGE_SIZE.
>
> Changes in V5:
> - Drop the 4k limit for normal readahead. (Jan Kara)
>
> Changes in V4:
> - Check for total node memory to decide whether we don't
> have local memory (jan Kara)
> - Add 4k page limit on readahead for normal and remote readahead (Linus)
> (Linus suggestion was 16MB limit).
>
> Changes in V3:
> - Drop iterating over numa nodes that calculates total free pages (Linus)
>
> Agree that we do not have control on allocation for readahead on a
> particular numa node and hence for remote readahead we can not further
> sanitize based on potential free pages of that node. and also we do
> not want to itererate through all nodes to find total free pages.
>
> Suggestions and comments welcome
> mm/readahead.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/mm/readahead.c b/mm/readahead.c
> index 0de2360..1fa0d6f 100644
> --- a/mm/readahead.c
> +++ b/mm/readahead.c
> @@ -233,14 +233,14 @@ int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
> return 0;
> }
>
> +#define MAX_READAHEAD ((512*4096)/PAGE_CACHE_SIZE)
> /*
> * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
> * sensible upper limit.
> */
> unsigned long max_sane_readahead(unsigned long nr)
> {
> - return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
> - + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
> + return min(nr, MAX_READAHEAD);
> }
>
> /*
> --
> 1.7.11.7
>
--
Jan Kara <[email protected]>
SUSE Labs, CR
On 02/18/2014 03:19 PM, Jan Kara wrote:
> On Tue 18-02-14 12:55:38, Raghavendra K T wrote:
>> Currently max_sane_readahead() returns zero on the cpu having no local memory node
>> which leads to readahead failure. Fix the readahead failure by returning
>> minimum of (requested pages, 512). Users running application on a memory-less cpu
>> which needs readahead such as streaming application see considerable boost in the
>> performance.
>>
>> Result:
>> fadvise experiment with FADV_WILLNEED on a PPC machine having memoryless CPU
>> with 1GB testfile ( 12 iterations) yielded around 46.66% improvement.
>>
>> fadvise experiment with FADV_WILLNEED on a x240 machine with 1GB testfile
>> 32GB* 4G RAM numa machine ( 12 iterations) showed no impact on the normal
>> NUMA cases w/ patch.
> Can you try one more thing please? Compare startup time of some big
> executable (Firefox or LibreOffice come to my mind) for the patched and
> normal kernel on a machine which wasn't hit by this NUMA issue. And don't
> forget to do "echo 3 >/proc/sys/vm/drop_caches" before each test to flush
> the caches. If this doesn't show significant differences, I'm OK with the
> patch.
>
Thanks Honza, I checked with firefox (starting to particular point)..
I do not see any difference. Both the case took around 14sec.
( some time it is even faster.. may be because we do not do free page
calculation?. )
On Tue 18-02-14 17:34:54, Raghavendra K T wrote:
> On 02/18/2014 03:19 PM, Jan Kara wrote:
> >On Tue 18-02-14 12:55:38, Raghavendra K T wrote:
> >>Currently max_sane_readahead() returns zero on the cpu having no local memory node
> >>which leads to readahead failure. Fix the readahead failure by returning
> >>minimum of (requested pages, 512). Users running application on a memory-less cpu
> >>which needs readahead such as streaming application see considerable boost in the
> >>performance.
> >>
> >>Result:
> >>fadvise experiment with FADV_WILLNEED on a PPC machine having memoryless CPU
> >>with 1GB testfile ( 12 iterations) yielded around 46.66% improvement.
> >>
> >>fadvise experiment with FADV_WILLNEED on a x240 machine with 1GB testfile
> >>32GB* 4G RAM numa machine ( 12 iterations) showed no impact on the normal
> >>NUMA cases w/ patch.
> > Can you try one more thing please? Compare startup time of some big
> >executable (Firefox or LibreOffice come to my mind) for the patched and
> >normal kernel on a machine which wasn't hit by this NUMA issue. And don't
> >forget to do "echo 3 >/proc/sys/vm/drop_caches" before each test to flush
> >the caches. If this doesn't show significant differences, I'm OK with the
> >patch.
> >
>
> Thanks Honza, I checked with firefox (starting to particular point)..
> I do not see any difference. Both the case took around 14sec.
Good. You can add my:
Acked-by: Jan Kara <[email protected]>
> ( some time it is even faster.. may be because we do not do free
> page calculation?. )
Hardly, that calculation is just a tiny amount of CPU time in the
startup of the application. If there is really a significant difference, it
might be because we don't preload stuff which isn't used in the end.
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
On Tue, 18 Feb 2014, Raghavendra K T wrote:
> Currently max_sane_readahead() returns zero on the cpu having no local memory node
> which leads to readahead failure. Fix the readahead failure by returning
> minimum of (requested pages, 512). Users running application on a memory-less cpu
> which needs readahead such as streaming application see considerable boost in the
> performance.
>
> Result:
> fadvise experiment with FADV_WILLNEED on a PPC machine having memoryless CPU
> with 1GB testfile ( 12 iterations) yielded around 46.66% improvement.
>
> fadvise experiment with FADV_WILLNEED on a x240 machine with 1GB testfile
> 32GB* 4G RAM numa machine ( 12 iterations) showed no impact on the normal
> NUMA cases w/ patch.
>
> Kernel Avg Stddev
> base 7.4975 3.92%
> patched 7.4174 3.26%
>
> Suggested-by: Linus Torvalds <[email protected]>
> [Andrew: making return value PAGE_SIZE independent]
> Signed-off-by: Raghavendra K T <[email protected]>
So this replaces
mm-readaheadc-fix-readahead-fail-for-no-local-memory-and-limit-readahead-pages.patch
in -mm correct?
> ---
> I would like to thank Honza, David for their valuable suggestions and
> patiently reviewing the patches.
>
> Changes in V6:
> - Just limit the readahead to 2MB on 4k pages system as suggested by Linus.
> and make it independent of PAGE_SIZE.
>
I'm not sure I understand why we want to be independent of PAGE_SIZE since
we're still relying on PAGE_CACHE_SIZE. Don't you mean to do
#define MAX_READAHEAD ((512*PAGE_SIZE)/PAGE_CACHE_SIZE)
instead?
On Tue, 18 Feb 2014 14:23:44 -0800 (PST) David Rientjes <[email protected]> wrote:
> On Tue, 18 Feb 2014, Raghavendra K T wrote:
>
> > Currently max_sane_readahead() returns zero on the cpu having no local memory node
> > which leads to readahead failure. Fix the readahead failure by returning
> > minimum of (requested pages, 512). Users running application on a memory-less cpu
> > which needs readahead such as streaming application see considerable boost in the
> > performance.
> >
> > Result:
> > fadvise experiment with FADV_WILLNEED on a PPC machine having memoryless CPU
> > with 1GB testfile ( 12 iterations) yielded around 46.66% improvement.
> >
> > fadvise experiment with FADV_WILLNEED on a x240 machine with 1GB testfile
> > 32GB* 4G RAM numa machine ( 12 iterations) showed no impact on the normal
> > NUMA cases w/ patch.
> >
> > Kernel Avg Stddev
> > base 7.4975 3.92%
> > patched 7.4174 3.26%
> >
> > Suggested-by: Linus Torvalds <[email protected]>
> > [Andrew: making return value PAGE_SIZE independent]
> > Signed-off-by: Raghavendra K T <[email protected]>
>
> So this replaces
> mm-readaheadc-fix-readahead-fail-for-no-local-memory-and-limit-readahead-pages.patch
> in -mm correct?
yup.
> > Changes in V6:
> > - Just limit the readahead to 2MB on 4k pages system as suggested by Linus.
> > and make it independent of PAGE_SIZE.
> >
>
> I'm not sure I understand why we want to be independent of PAGE_SIZE since
> we're still relying on PAGE_CACHE_SIZE. Don't you mean to do
>
> #define MAX_READAHEAD ((512*PAGE_SIZE)/PAGE_CACHE_SIZE)
MAX_READAHEAD is in units of "pages".
This:
+#define MAX_READAHEAD ((512*4096)/PAGE_CACHE_SIZE)
means "two megabytes", and is implemented in a way to ensure that
MAX_READAHEAD=2mb on 4k pagesize as well as on 64k pagesize. Because
we don't want variations in PAGE_SIZE to cause alterations in readahead
behavior.
On Tue, 18 Feb 2014, Andrew Morton wrote:
> > I'm not sure I understand why we want to be independent of PAGE_SIZE since
> > we're still relying on PAGE_CACHE_SIZE. Don't you mean to do
> >
> > #define MAX_READAHEAD ((512*PAGE_SIZE)/PAGE_CACHE_SIZE)
>
> MAX_READAHEAD is in units of "pages".
>
> This:
>
> +#define MAX_READAHEAD ((512*4096)/PAGE_CACHE_SIZE)
>
> means "two megabytes", and is implemented in a way to ensure that
> MAX_READAHEAD=2mb on 4k pagesize as well as on 64k pagesize. Because
> we don't want variations in PAGE_SIZE to cause alterations in readahead
> behavior.
>
Ah, ok, so 2MB is the magic value that we limit readhead to on all
architectures. 512 * 4096 is a strange way to write 2MB, but ok :)