mm thp: shrink deferred split THPs harder
Subsystems (2)
PATCH 1/1 mm thp: shrink deferred split THPs harder
The deferred split THPs may get accumulated with some workloads, they would get shrunk when memory pressure is hit. Now we use DEFAULT_SEEKS to determine how many objects would get scanned then split if possible, but actually they are not like other system cache objects, i.e. inode cache which would incur extra I/O if over reclaimed, the unmapped pages will not be accessed anymore, so we could shrink them more aggressively.
We could shrink THPs more pro-actively even though memory pressure is not hit, however, IMHO waiting for memory pressure is still a good compromise and trade-off. And, we do have simpler ways to shrink these objects harder until we have to take other means do pro-actively drain.
Change shrinker->seeks to 0 to shrink deferred split THPs harder.
--- mm/huge_memory.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
mm/huge_memory.c +1 -1
@@ -2955,7 +2955,7 @@ static unsigned long deferred_split_scan(struct shrinker *shrink,static struct shrinker deferred_split_shrinker = { .count_objects = deferred_split_count, .scan_objects = deferred_split_scan, - .seeks =DEFAULT_SEEKS , + .seeks =0 , .flags =SHRINKER_NUMA_AWARE |SHRINKER_MEMCG_AWARE |SHRINKER_NONSLAB , };
Signature · 1 line
1.8.3.1
Quoted 7 lines · Yang Shi
The deferred split THPs may get accumulated with some workloads, they would get shrunk when memory pressure is hit. Now we use
DEFAULT_SEEKSto determine how many objects would get scanned then split if possible, but actually they are not like other system cache objects, i.e. inode cache which would incur extra I/O if over reclaimed, the unmapped pages will not be accessed anymore, so we could shrink them more aggressively.
Quoted 6 lines
We could shrink THPs more pro-actively even though memory pressure is not hit, however, IMHO waiting for memory pressure is still a good compromise and trade-off. And, we do have simpler ways to shrink these objects harder until we have to take other means do pro-actively drain.
Change
shrinker->seeksto 0 to shrink deferred split THPs harder.
Do you have any numbers on the effect of this patch.
Btw. the whole thing is getting more and more complex and I still believe the approach is just wrong. We are tunning for something that doesn't really belong to the memory reclaim in the first place IMHO.
Quoted 27 lines
Cc Kirill A. Shutemov <[email protected]>Cc Kirill Tkhai <[email protected]>Cc Johannes Weiner <[email protected]>Cc Michal Hocko <[email protected]>Cc Hugh Dickins <[email protected]>Cc Shakeel Butt <[email protected]>Cc David Rientjes <[email protected]>Signed-off-by Yang Shi <[email protected]>--- mm/huge_memory.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)mm/huge_memory.c +1
-1@@ -2955,7 +2955,7 @@ static unsigned long deferred_split_scan(struct shrinker *shrink,static struct shrinker deferred_split_shrinker = { .count_objects = deferred_split_count, .scan_objects = deferred_split_scan, - .seeks =DEFAULT_SEEKS , + .seeks =0 , .flags =SHRINKER_NUMA_AWARE |SHRINKER_MEMCG_AWARE |SHRINKER_NONSLAB , };Signature · 1 line
1.8.3.1
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 8 lines · Michal Hocko
Quoted 7 lines · Yang Shi
The deferred split THPs may get accumulated with some workloads, they would get shrunk when memory pressure is hit. Now we use
DEFAULT_SEEKSto determine how many objects would get scanned then split if possible, but actually they are not like other system cache objects, i.e. inode cache which would incur extra I/O if over reclaimed, the unmapped pages will not be accessed anymore, so we could shrink them more aggressively.
Quoted 7 lines
Quoted 6 lines
We could shrink THPs more pro-actively even though memory pressure is not hit, however, IMHO waiting for memory pressure is still a good compromise and trade-off. And, we do have simpler ways to shrink these objects harder until we have to take other means do pro-actively drain.
Change
shrinker->seeksto 0 to shrink deferred split THPs harder.Do you have any numbers on the effect of this patch.
Yes, this patch would make THPs get split earlier.
For example, I have a test case which generates around 4G deferred split THPs (2K huge pages). With the default seeks, THPs would start to get split when priority reaches 6 since nr_to_scan depends on priority and shrinker->seeks. With this patch it would start to get split at the very beginning (priority 12).
IMHO, somehow this would achieve the similar effect with pro-actively draining.
Quoted 4 lines
Btw. the whole thing is getting more and more complex and I still believe the approach is just wrong. We are tunning for something that doesn't really belong to the memory reclaim in the first place IMHO.
Maybe, but it is not clear to me that other approaches would be universally better than the current one unless we could split the page right away.
Quoted 28 lines
Quoted 28 lines
Cc Kirill A. Shutemov <[email protected]>Cc Kirill Tkhai <[email protected]>Cc Johannes Weiner <[email protected]>Cc Michal Hocko <[email protected]>Cc Hugh Dickins <[email protected]>Cc Shakeel Butt <[email protected]>Cc David Rientjes <[email protected]>Signed-off-by Yang Shi <[email protected]>--- mm/huge_memory.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)mm/huge_memory.c +1
-1@@ -2955,7 +2955,7 @@ static unsigned long deferred_split_scan(struct shrinker *shrink,static struct shrinker deferred_split_shrinker = { .count_objects = deferred_split_count, .scan_objects = deferred_split_scan, - .seeks =DEFAULT_SEEKS , + .seeks =0 , .flags =SHRINKER_NUMA_AWARE |SHRINKER_MEMCG_AWARE |SHRINKER_NONSLAB , };Signature · 1 line
1.8.3.1