Against: 2.6.31-rc6
We noticed very erratic behavior [throughput] with the AIM7 shared
workload running on recent distro [SLES11] and mainline kernels on
an 8-socket, 32-core, 256GB x86_64 platform. On the SLES11 kernel
[2.6.27.19+] with Barcelona processors, as we increased the load
[10s of thousands of tasks], the throughput would vary between two
"plateaus"--one at ~65K jobs per minute and one at ~130K jpm. The
simple patch below causes the results to smooth out at the ~130k
plateau.
But wait, there's more:
We do not see this behavior on smaller platforms--e.g., 4 socket/8
core. This could be the result of the larger number of cpus on
the larger platform--a scalability issue--or it could be the result
of the larger number of interconnect "hops" between some nodes in
this platform and how the tasks for a given load end up distributed
over the nodes' cpus and memories--a stochastic NUMA effect.
The variability in the results are less pronounced [on the same
platform] with Shanghai processors and with mainline kernels. With
31-rc6 on Shanghai processors and 288 file systems on 288 fibre
attached storage volumes, the curves [jpm vs load] are both quite
flat with the patched kernel consistently producing ~3.9% better
throughput [~80K jpm vs ~77K jpm] than the unpatched kernel.
Profiling indicated that the "slow" runs were incurring high[er]
contention on an anon_vma lock in vma_adjust(), apparently called
from the sbrk() system call.
The patch:
A comment in mm/mmap.c:vma_adjust() suggests that we don't really
need the anon_vma lock when we're only adjusting the end of a vma,
as is the case for brk(). The comment questions whether it's worth
while to optimize for this case. Apparently, on the newer, larger
x86_64 platforms, with interesting NUMA topologies, it is worth
while--especially considering that the patch [if correct!] is
quite simple.
We can detect this condition--no overlap with next vma--by noting
a NULL "importer". The anon_vma pointer will also be NULL in this
case, so simply avoid loading vma->anon_vma to avoid the lock.
However, we apparently DO need to take the anon_vma lock when
we're inserting a vma ['insert' non-NULL] even when we have no
overlap [NULL "importer"], so we need to check for 'insert', as well.
I have tested with and without the 'file || ' test in the patch.
This does not seem to matter for stability nor performance. I
left this check/filter in, so we only optimize away the
anon_vma lock acquisition when adjusting the end of a non-
importing, non-inserting, anon vma.
If accepted, this patch may benefit the stable tree as well.
Signed-off-by: Lee Schermerhorn <[email protected]>
mm/mmap.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
Index: linux-2.6.31-rc6/mm/mmap.c
===================================================================
--- linux-2.6.31-rc6.orig/mm/mmap.c 2009-08-19 14:34:13.000000000 -0400
+++ linux-2.6.31-rc6/mm/mmap.c 2009-08-19 14:53:24.000000000 -0400
@@ -573,9 +573,9 @@ again: remove_next = 1 + (end > next->
/*
* When changing only vma->vm_end, we don't really need
- * anon_vma lock: but is that case worth optimizing out?
+ * anon_vma lock.
*/
- if (vma->anon_vma)
+ if ((file || insert || importer) && vma->anon_vma)
anon_vma = vma->anon_vma;
if (anon_vma) {
spin_lock(&anon_vma->lock);
[I've kept [email protected] in the Cc list only to put on record
that I don't think that this patch is really suitable for -stable]
On Wed, 9 Sep 2009, Lee Schermerhorn wrote:
>
Thanks for this. Interesting stuff snipped and repeated below.
>
> A comment in mm/mmap.c:vma_adjust() suggests that we don't really
> need the anon_vma lock when we're only adjusting the end of a vma,
I feel a warm smugness from having foreseen the possibility that we
might want to optimize that away. I didn't do so, because it needs
more thought (and more branches) than seemed worthwhile at the time:
you've found similar difficulty, but now it does seem worthwhile.
> We can detect this condition--no overlap with next vma--by noting
> a NULL "importer". The anon_vma pointer will also be NULL in this
> case, so simply avoid loading vma->anon_vma to avoid the lock.
> However, we apparently DO need to take the anon_vma lock when
> we're inserting a vma ['insert' non-NULL] even when we have no
> overlap [NULL "importer"], so we need to check for 'insert', as well.
Those importer and insert checks are good and relevant, but not
quite enough. The anon_vma lock should also be guaranteeing the
integrity of the relationship between vm_start and vm_pgoff for
all the vmas attached to the anon_vma, so that rmap.c can rely
upon vma_address() to work correctly while it holds anon_vma lock.
That's a considerably less important consideration than the integrity
of the list threading itself. Anything BUGging on a wrong page->index
is holding mmap_sem, which would keep vma_adjust off. So it's just a
matter of whether rmap.c can be expected to find all instances of a
page at all times, which nothing absolutely requires (and in checking
this patch, I notice fs/exec.c's shift_arg_pages() use of vma_adjust()
a little violatory in that respect).
But it is something the anon_vma lock has protected in the past,
and it shouldn't affect your sbrk() case at all, so I'd like to
check we're not changing vm_start too (vm_pgoff should be changing
with it, but shift_arg_pages() deals with that in a different way,
keeping vm_pgoff unchanged but shifting the pages).
(Compare with how stack's expand_downwards() has anon_vma_lock()
when it adjusts vm_start and vm_pgoff - though that's also because
it has only down_read of mmap_sem, not the down_write we'd usually
require for such adjustments.)
>
> I have tested with and without the 'file || ' test in the patch.
> This does not seem to matter for stability nor performance. I
> left this check/filter in, so we only optimize away the
> anon_vma lock acquisition when adjusting the end of a non-
> importing, non-inserting, anon vma.
I dislike that: the "file" test just has no relevance at all
(beyond that you've no interest in the case when there is a file),
and two years down the line will make people like me worry for
hours on end what it's there for. I've removed it below.
>
> If accepted, this patch may benefit the stable tree as well.
We seem to have a different perception of what the stable tree is for!
But I'm not against any distro picking up this patch if it chooses.
Here's a version with my signoff:
[PATCH] mmap: avoid unnecessary anon_vma lock acquisition in vma_adjust()
From: Lee Schermerhorn <[email protected]>
We noticed very erratic behavior [throughput] with the AIM7 shared
workload running on recent distro [SLES11] and mainline kernels on
an 8-socket, 32-core, 256GB x86_64 platform. On the SLES11 kernel
[2.6.27.19+] with Barcelona processors, as we increased the load
[10s of thousands of tasks], the throughput would vary between two
"plateaus"--one at ~65K jobs per minute and one at ~130K jpm. The
simple patch below causes the results to smooth out at the ~130k
plateau.
But wait, there's more:
We do not see this behavior on smaller platforms--e.g., 4 socket/8
core. This could be the result of the larger number of cpus on
the larger platform--a scalability issue--or it could be the result
of the larger number of interconnect "hops" between some nodes in
this platform and how the tasks for a given load end up distributed
over the nodes' cpus and memories--a stochastic NUMA effect.
The variability in the results are less pronounced [on the same
platform] with Shanghai processors and with mainline kernels. With
31-rc6 on Shanghai processors and 288 file systems on 288 fibre
attached storage volumes, the curves [jpm vs load] are both quite
flat with the patched kernel consistently producing ~3.9% better
throughput [~80K jpm vs ~77K jpm] than the unpatched kernel.
Profiling indicated that the "slow" runs were incurring high[er]
contention on an anon_vma lock in vma_adjust(), apparently called
from the sbrk() system call.
The patch:
A comment in mm/mmap.c:vma_adjust() suggests that we don't really
need the anon_vma lock when we're only adjusting the end of a vma,
as is the case for brk(). The comment questions whether it's worth
while to optimize for this case. Apparently, on the newer, larger
x86_64 platforms, with interesting NUMA topologies, it is worth
while--especially considering that the patch [if correct!] is
quite simple.
We can detect this condition--no overlap with next vma--by noting
a NULL "importer". The anon_vma pointer will also be NULL in this
case, so simply avoid loading vma->anon_vma to avoid the lock.
However, we DO need to take the anon_vma lock when we're inserting a
vma ['insert' non-NULL] even when we have no overlap [NULL "importer"],
so we need to check for 'insert', as well. And Hugh points out that
we should also take it when adjusting vm_start (so that rmap.c can
rely upon vma_address() while it holds the anon_vma lock).
Signed-off-by: Lee Schermerhorn <[email protected]>
Signed-off-by: Hugh Dickins <[email protected]>
Cc: Nick Piggin <[email protected]>
Cc: Eric Whitney <[email protected]>
---
mm/mmap.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
--- 2.6.31/mm/mmap.c 2009-09-09 23:13:59.000000000 +0100
+++ linux/mm/mmap.c 2009-09-13 13:08:40.000000000 +0100
@@ -570,9 +570,9 @@ again: remove_next = 1 + (end > next->
/*
* When changing only vma->vm_end, we don't really need
- * anon_vma lock: but is that case worth optimizing out?
+ * anon_vma lock.
*/
- if (vma->anon_vma)
+ if (vma->anon_vma && (insert || importer || start != vma->vm_start))
anon_vma = vma->anon_vma;
if (anon_vma) {
spin_lock(&anon_vma->lock);
On Sun, 2009-09-13 at 14:24 +0100, Hugh Dickins wrote:
> [I've kept [email protected] in the Cc list only to put on record
> that I don't think that this patch is really suitable for -stable]
>
> On Wed, 9 Sep 2009, Lee Schermerhorn wrote:
> >
> Thanks for this. Interesting stuff snipped and repeated below.
Hugh:
Thank you for reviewing and correcting the patch.
Andrew:
Shall I resend? Or, can you grab Hugh's patch from his message?
Lee
>
> >
> > A comment in mm/mmap.c:vma_adjust() suggests that we don't really
> > need the anon_vma lock when we're only adjusting the end of a vma,
>
> I feel a warm smugness from having foreseen the possibility that we
> might want to optimize that away. I didn't do so, because it needs
> more thought (and more branches) than seemed worthwhile at the time:
> you've found similar difficulty, but now it does seem worthwhile.
>
> > We can detect this condition--no overlap with next vma--by noting
> > a NULL "importer". The anon_vma pointer will also be NULL in this
> > case, so simply avoid loading vma->anon_vma to avoid the lock.
> > However, we apparently DO need to take the anon_vma lock when
> > we're inserting a vma ['insert' non-NULL] even when we have no
> > overlap [NULL "importer"], so we need to check for 'insert', as well.
>
> Those importer and insert checks are good and relevant, but not
> quite enough. The anon_vma lock should also be guaranteeing the
> integrity of the relationship between vm_start and vm_pgoff for
> all the vmas attached to the anon_vma, so that rmap.c can rely
> upon vma_address() to work correctly while it holds anon_vma lock.
>
> That's a considerably less important consideration than the integrity
> of the list threading itself. Anything BUGging on a wrong page->index
> is holding mmap_sem, which would keep vma_adjust off. So it's just a
> matter of whether rmap.c can be expected to find all instances of a
> page at all times, which nothing absolutely requires (and in checking
> this patch, I notice fs/exec.c's shift_arg_pages() use of vma_adjust()
> a little violatory in that respect).
>
> But it is something the anon_vma lock has protected in the past,
> and it shouldn't affect your sbrk() case at all, so I'd like to
> check we're not changing vm_start too (vm_pgoff should be changing
> with it, but shift_arg_pages() deals with that in a different way,
> keeping vm_pgoff unchanged but shifting the pages).
>
> (Compare with how stack's expand_downwards() has anon_vma_lock()
> when it adjusts vm_start and vm_pgoff - though that's also because
> it has only down_read of mmap_sem, not the down_write we'd usually
> require for such adjustments.)
>
> >
> > I have tested with and without the 'file || ' test in the patch.
> > This does not seem to matter for stability nor performance. I
> > left this check/filter in, so we only optimize away the
> > anon_vma lock acquisition when adjusting the end of a non-
> > importing, non-inserting, anon vma.
>
> I dislike that: the "file" test just has no relevance at all
> (beyond that you've no interest in the case when there is a file),
> and two years down the line will make people like me worry for
> hours on end what it's there for. I've removed it below.
>
> >
> > If accepted, this patch may benefit the stable tree as well.
>
> We seem to have a different perception of what the stable tree is for!
> But I'm not against any distro picking up this patch if it chooses.
> Here's a version with my signoff:
>
>
> [PATCH] mmap: avoid unnecessary anon_vma lock acquisition in vma_adjust()
>
> From: Lee Schermerhorn <[email protected]>
>
> We noticed very erratic behavior [throughput] with the AIM7 shared
> workload running on recent distro [SLES11] and mainline kernels on
> an 8-socket, 32-core, 256GB x86_64 platform. On the SLES11 kernel
> [2.6.27.19+] with Barcelona processors, as we increased the load
> [10s of thousands of tasks], the throughput would vary between two
> "plateaus"--one at ~65K jobs per minute and one at ~130K jpm. The
> simple patch below causes the results to smooth out at the ~130k
> plateau.
>
> But wait, there's more:
>
> We do not see this behavior on smaller platforms--e.g., 4 socket/8
> core. This could be the result of the larger number of cpus on
> the larger platform--a scalability issue--or it could be the result
> of the larger number of interconnect "hops" between some nodes in
> this platform and how the tasks for a given load end up distributed
> over the nodes' cpus and memories--a stochastic NUMA effect.
>
> The variability in the results are less pronounced [on the same
> platform] with Shanghai processors and with mainline kernels. With
> 31-rc6 on Shanghai processors and 288 file systems on 288 fibre
> attached storage volumes, the curves [jpm vs load] are both quite
> flat with the patched kernel consistently producing ~3.9% better
> throughput [~80K jpm vs ~77K jpm] than the unpatched kernel.
>
> Profiling indicated that the "slow" runs were incurring high[er]
> contention on an anon_vma lock in vma_adjust(), apparently called
> from the sbrk() system call.
>
> The patch:
>
> A comment in mm/mmap.c:vma_adjust() suggests that we don't really
> need the anon_vma lock when we're only adjusting the end of a vma,
> as is the case for brk(). The comment questions whether it's worth
> while to optimize for this case. Apparently, on the newer, larger
> x86_64 platforms, with interesting NUMA topologies, it is worth
> while--especially considering that the patch [if correct!] is
> quite simple.
>
> We can detect this condition--no overlap with next vma--by noting
> a NULL "importer". The anon_vma pointer will also be NULL in this
> case, so simply avoid loading vma->anon_vma to avoid the lock.
>
> However, we DO need to take the anon_vma lock when we're inserting a
> vma ['insert' non-NULL] even when we have no overlap [NULL "importer"],
> so we need to check for 'insert', as well. And Hugh points out that
> we should also take it when adjusting vm_start (so that rmap.c can
> rely upon vma_address() while it holds the anon_vma lock).
>
> Signed-off-by: Lee Schermerhorn <[email protected]>
> Signed-off-by: Hugh Dickins <[email protected]>
> Cc: Nick Piggin <[email protected]>
> Cc: Eric Whitney <[email protected]>
> ---
>
> mm/mmap.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> --- 2.6.31/mm/mmap.c 2009-09-09 23:13:59.000000000 +0100
> +++ linux/mm/mmap.c 2009-09-13 13:08:40.000000000 +0100
> @@ -570,9 +570,9 @@ again: remove_next = 1 + (end > next->
>
> /*
> * When changing only vma->vm_end, we don't really need
> - * anon_vma lock: but is that case worth optimizing out?
> + * anon_vma lock.
> */
> - if (vma->anon_vma)
> + if (vma->anon_vma && (insert || importer || start != vma->vm_start))
> anon_vma = vma->anon_vma;
> if (anon_vma) {
> spin_lock(&anon_vma->lock);
>
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to [email protected]. For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"[email protected]"> [email protected] </a>