lkml.org 
[lkml]   [2011]   [Dec]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: [PATCH 1/3] slub: set a criteria for slub node partial adding
From
Date
On Fri, 2011-12-02 at 22:43 +0800, Christoph Lameter wrote:
> On Fri, 2 Dec 2011, Alex Shi wrote:
>
> > From: Alex Shi <alexs@intel.com>
> >
> > Times performance regression were due to slub add to node partial head
> > or tail. That inspired me to do tunning on the node partial adding, to
> > set a criteria for head or tail position selection when do partial
> > adding.
> > My experiment show, when used objects is less than 1/4 total objects
> > of slub performance will get about 1.5% improvement on netperf loopback
> > testing with 2048 clients, wherever on our 4 or 2 sockets platforms,
> > includes sandbridge or core2.
>
> The number of free objects in a slab may have nothing to do with cache
> hotness of all objects in the slab. You can only be sure that one object
> (the one that was freed) is cache hot. Netperf may use them in sequence
> and therefore you are likely to get series of frees on the same slab
> page. How are other benchmarks affected by this change?


I did some experiments on add_partial judgment against rc4, like to put
the slub into node partial head or tail according to free objects, or
like Eric's suggest to combine the external parameter, like below:

n->nr_partial++;
- if (tail == DEACTIVATE_TO_TAIL)
+ if (tail == DEACTIVATE_TO_TAIL ||
+ page->inuse > page->objects /2)
list_add_tail(&page->lru, &n->partial);
else
list_add(&page->lru, &n->partial);
But the result is out of my expectation before. Now we set all of slub
into the tail of node partial, we get the best performance, even it is
just a slight improvement.

{
n->nr_partial++;
- if (tail == DEACTIVATE_TO_TAIL)
- list_add_tail(&page->lru, &n->partial);
- else
- list_add(&page->lru, &n->partial);
+ list_add_tail(&page->lru, &n->partial);
}
This change can bring about 2% improvement on our WSM-ep machine, and 1%
improvement on our SNB-ep and NHM-ex machine. and no clear effect for
core2 machine. on hackbench process benchmark.

./hackbench 100 process 2000

For multiple clients loopback netperf, only a suspicious 1% improvement
on our 2 sockets machine. and others have no clear effect.

But, when I check the deactivate_to_head/to_tail statistics on original
code, the to_head is just hundreds or thousands times, while to_tail is
called about teens millions times.

David, could you like to try above change? move all slub to partial
tail.

add_partial statistics collection patch:
---
commit 1ff731282acb521f3a7c2e3fb94d35ec4d0ff07e
Author: Alex Shi <alex.shi@intel.com>
Date: Fri Dec 9 18:12:14 2011 +0800
slub: statistics collection for add_partial

diff --git a/mm/slub.c b/mm/slub.c
index 5843846..a2b1143 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1904,10 +1904,11 @@ static void unfreeze_partials(struct kmem_cache *s)
if (l != m) {
if (l == M_PARTIAL)
remove_partial(n, page);
- else
+ else {
add_partial(n, page,
DEACTIVATE_TO_TAIL);
-
+ stat(s, DEACTIVATE_TO_TAIL);
+ }
l = m;
}

@@ -2480,6 +2481,7 @@ static void __slab_free(struct kmem_cache *s, struct page *page,
remove_full(s, page);
add_partial(n, page, DEACTIVATE_TO_TAIL);
stat(s, FREE_ADD_PARTIAL);
+ stat(s, DEACTIVATE_TO_TAIL);
}
}
spin_unlock_irqrestore(&n->list_lock, flags);




\
 
 \ /
  Last update: 2011-12-09 09:35    [from the cache]
©2003-2011 Jasper Spaans