>> > For example, in a 4 socket 64 CPU system, create a sparse file and 64 processes,
>> > processes shared map to the file. Each process read access the whole file and
>> > then exit. The process exit will do unmap_vmas() and cause a lot of
>> > activate_page() call. In such workload, we saw about 58% total time reduction
>> > with below patch.
>>
>> I'm not sure this. Why process exiting on your workload call unmap_vmas?
>
> Trick question?
>
> Getting rid of a mm on process exit requires unmapping the vmas.
Oops, I misparsed unmap_vmas() and unuse_vma(). thanks fix me.
Ho Hum, zap_pte_range() call mark_page_accessed(). it was introduced slightly
recent (below).
----------------------------------------------------------------------------------
commit bf3f3bc5e734706730c12a323f9b2068052aa1f0
Author: Nick Piggin <npiggin@xxxxxxx>
Date: Tue Jan 6 14:38:55 2009 -0800
mm: don't mark_page_accessed in fault path
----------------------------------------------------------------------------------
So, I guess we can apply following small optimization.
the intention is, if the exiting process is last mapped one,
we don't need to refrect its pte_young() because it is good
sign that the page is never touched.
Of cource, this is offtopic. On Shaouhua's testcase, 64 processes
shared map to the file.
-------------------------------------------------------------------------------
diff --git a/mm/memory.c b/mm/memory.c
index 833952d..ebdfcb1 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -951,7 +951,8 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,
if (pte_dirty(ptent))
set_page_dirty(page);
if (pte_young(ptent) &&
- likely(!VM_SequentialReadHint(vma)))
+ (page_mapcount(page) != 1) &&
+ !VM_SequentialReadHint(vma))
mark_page_accessed(page);
rss[MM_FILEPAGES]--;
-------------------------------------------------------------------------------
One more offtopic:
we need to move ClearPageReferenced() from mark_page_accessed()
to __activate_page() honorly. unuse_vma() case also need to clear
page-referenced bit. I think.
Anyway, I agree original patch concept is fine. Thanks a lot of productive
information!
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>