* Ingo Molnar <mingo@xxxxxxxxxx> wrote:
> 1)
>
> This patch might solve the remapping
> (remove_migration_ptes()), but does not solve the anon-vma
> locking done in the first, unmapping step of pte-migration -
> which is done via try_to_unmap(): which is a generic VM
> function used by swapout too, so callers do not necessarily
> hold the mmap_sem.
>
> A new TTU flag might solve it although I detest flag-driven
> locking semantics with a passion:
>
> Splitting out unlocked versions of try_to_unmap_anon(),
> try_to_unmap_ksm(), try_to_unmap_file() and constructing an
> unlocked try_to_unmap() out of them, to be used by the
> migration code, would be the cleaner option.
So as a quick concept hack I wrote the patch attached below.
(It's not signed off, see the patch description text for the
reason.)
With this applied I get the same good 4x JVM performance:
spec1.txt: throughput = 157471.10 SPECjbb2005 bops
spec2.txt: throughput = 157817.09 SPECjbb2005 bops
spec3.txt: throughput = 157581.79 SPECjbb2005 bops
spec4.txt: throughput = 157890.26 SPECjbb2005 bops
--------------------------
SUM: throughput = 630760.24 SPECjbb2005 bops
... because the JVM workload did not trigger the migration
scalability threshold to begin with.
Mainline 4xJVM SPECjbb performance:
spec1.txt: throughput = 128575.47 SPECjbb2005 bops
spec2.txt: throughput = 125767.24 SPECjbb2005 bops
spec3.txt: throughput = 130042.30 SPECjbb2005 bops
spec4.txt: throughput = 128155.32 SPECjbb2005 bops
--------------------------
SUM: throughput = 512540.33 SPECjbb2005 bops
# (32 CPUs, 4 instances, 8 warehouses each, 240 seconds runtime, !THP)
But !THP/4K numa02 performance went trough the roof!
Mainline !THP numa02 performance:
40.918 secs runtime/thread
26.051 secs fastest (min) thread time
59.229 secs elapsed (max) thread time [ spread: -28.0% ]
26.844 GB data processed, per thread
858.993 GB data processed, total
2.206 nsecs/byte/thread
0.453 GB/sec/thread
14.503 GB/sec total
numa/core v18 + migration-locking-enhancements, !THP:
18.543 secs runtime/thread
17.721 secs fastest (min) thread time
19.262 secs elapsed (max) thread time [ spread: -4.0% ]
26.844 GB data processed, per thread
858.993 GB data processed, total
0.718 nsecs/byte/thread
1.394 GB/sec/thread
44.595 GB/sec total
as you can see the performance of each of the 32 threads is
within a tight bound:
17.721 secs fastest (min) thread time
19.262 secs elapsed (max) thread time [ spread: -4.0% ]
... with very little spread between them.
So this is roughly as good as it can get without hard binding -
and according to my limited testing the numa02 workload is
20-30% faster than the AutoNUMA or balancenuma kernels on the
same hardware/kernel combo. The above numa02 result now also
gets reasonably close to the numa/core +THP numa02 numbers (to
within 10%).
As expected there's a lot of TLB flushing going on, but, and
this was unexpected to me, even maximally pushing the migration
code does not trigger anything pathological on this 4-node
system - so while the TLB optimization will be a welcome
enhancement, it's not a must-have at this stage.
I'll do a cleaner version of this patch and I'll test on a
larger system with a large NUMA factor too to make sure we don't
need the TLB optimization on !THP.
So I think (assuming that I have not overlooked something
critical in these patches!), with these two fixes all the
difficult known regressions in numa/core are fixed.
I'll do more testing with broader workloads and on more systems
to ascertain this.
Thanks,
Ingo
---------------->
Subject: mm/migration: Remove anon vma locking from try_to_unmap() use
From: Ingo Molnar <mingo@xxxxxxxxxx>
Date: Sat Dec 1 11:22:09 CET 2012
As outlined in:
mm/migration: Don't lock anon vmas in rmap_walk_anon()
the process-global anon vma mutex locking of the page migration
code can be very expensive.
This removes the second (and last) use of that mutex from the
migration code: try_to_unmap().
Since try_to_unmap() is used by swapout and filesystem code
as well, which does not hold the mmap_sem, we only want to
do this optimization from the migration path.
This patch is ugly and should be replaced via a
try_to_unmap_locked() variant instead which offers us the
unlocked codepath, but it's good enough for testing purposes.
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Cc: Peter Zijlstra <a.p.zijlstra@xxxxxxxxx>
Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
Cc: Rik van Riel <riel@xxxxxxxxxx>
Cc: Mel Gorman <mgorman@xxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: Hugh Dickins <hughd@xxxxxxxxxx>
Not-Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
---
include/linux/rmap.h | 2 +-
mm/huge_memory.c | 2 +-
mm/memory-failure.c | 2 +-
mm/rmap.c | 13 ++++++++++---
4 files changed, 13 insertions(+), 6 deletions(-)
Index: linux/include/linux/rmap.h
===================================================================
--- linux.orig/include/linux/rmap.h
+++ linux/include/linux/rmap.h
@@ -220,7 +220,7 @@ int try_to_munlock(struct page *);
/*
* Called by memory-failure.c to kill processes.
*/
-struct anon_vma *page_lock_anon_vma(struct page *page);
+struct anon_vma *page_lock_anon_vma(struct page *page, enum ttu_flags flags);
void page_unlock_anon_vma(struct anon_vma *anon_vma);
int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma);
Index: linux/mm/huge_memory.c
===================================================================
--- linux.orig/mm/huge_memory.c
+++ linux/mm/huge_memory.c
@@ -1645,7 +1645,7 @@ int split_huge_page(struct page *page)
int ret = 1;
BUG_ON(!PageAnon(page));
- anon_vma = page_lock_anon_vma(page);
+ anon_vma = page_lock_anon_vma(page, 0);
if (!anon_vma)
goto out;
ret = 0;
Index: linux/mm/memory-failure.c
===================================================================
--- linux.orig/mm/memory-failure.c
+++ linux/mm/memory-failure.c
@@ -402,7 +402,7 @@ static void collect_procs_anon(struct pa
struct anon_vma *av;
pgoff_t pgoff;
- av = page_lock_anon_vma(page);
+ av = page_lock_anon_vma(page, 0);
if (av == NULL) /* Not actually mapped anymore */
return;
Index: linux/mm/rmap.c
===================================================================
--- linux.orig/mm/rmap.c
+++ linux/mm/rmap.c
@@ -442,7 +442,7 @@ out:
* atomic op -- the trylock. If we fail the trylock, we fall back to getting a
* reference like with page_get_anon_vma() and then block on the mutex.
*/
-struct anon_vma *page_lock_anon_vma(struct page *page)
+struct anon_vma *page_lock_anon_vma(struct page *page, enum ttu_flags flags)
{
struct anon_vma *anon_vma = NULL;
struct anon_vma *root_anon_vma;
@@ -456,6 +456,13 @@ struct anon_vma *page_lock_anon_vma(stru
goto out;
anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
+ /*
+ * The migration code paths are already holding the mmap_sem,
+ * so the anon vma cannot go away from under us - return it:
+ */
+ if (flags & TTU_MIGRATION)
+ goto out;
+
root_anon_vma = ACCESS_ONCE(anon_vma->root);
if (mutex_trylock(&root_anon_vma->mutex)) {
/*
@@ -732,7 +739,7 @@ static int page_referenced_anon(struct p
struct anon_vma_chain *avc;
int referenced = 0;
- anon_vma = page_lock_anon_vma(page);
+ anon_vma = page_lock_anon_vma(page, 0);
if (!anon_vma)
return referenced;
@@ -1474,7 +1481,7 @@ static int try_to_unmap_anon(struct page
struct anon_vma_chain *avc;
int ret = SWAP_AGAIN;
- anon_vma = page_lock_anon_vma(page);
+ anon_vma = page_lock_anon_vma(page, flags);
if (!anon_vma)
return ret;
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>