From: Andrew Morton <akpm@linux-foundation.org>
To: Wei Yang <richard.weiyang@gmail.com>
Cc: cl@linux.com, penberg@kernel.org, mhocko@kernel.org,
linux-mm@kvack.org, Joonsoo Kim <iamjoonsoo.kim@lge.com>,
David Rientjes <rientjes@google.com>
Subject: Re: [PATCH v2] mm/slub: improve performance by skipping checked node in get_any_partial()
Date: Thu, 20 Dec 2018 14:41:07 -0800 [thread overview]
Message-ID: <20181220144107.9376344c2be687615ea9aa69@linux-foundation.org> (raw)
In-Reply-To: <20181120033119.30013-1-richard.weiyang@gmail.com>
Could someone please review this?
Thanks.
From: Wei Yang <richard.weiyang@gmail.com>
Subject: mm/slub.c: improve performance by skipping checked node in get_any_partial()
1. Background
Current slub has three layers:
* cpu_slab
* percpu_partial
* per node partial list
Slub allocator tries to get an object from top to bottom. When it
can't get an object from the upper two layers, it will search the per
node partial list. The is done in get_partial().
The abstraction of get_partial() look like this:
get_partial()
get_partial_node()
get_any_partial()
for_each_zone_zonelist()
The idea behind this is: first try a local node, then try other nodes
if caller doesn't specify a node.
2. Room for Improvement
When we look one step deeper in get_any_partial(), it tries to get a
proper node by for_each_zone_zonelist(), which iterates on the
node_zonelists.
This behavior would introduce some redundant check on the same node.
Because:
* the local node is already checked in get_partial_node()
* one node may have several zones on node_zonelists
3. Solution Proposed in Patch
We could reduce these redundant check by record the last unsuccessful
node and then skip it.
4. Tests & Result
After some tests, the result shows this may improve the system a little,
especially on a machine with only one node.
4.1 Test Description
There are two cases for two system configurations.
Test Cases:
1. counter comparison
2. kernel build test
System Configuration:
1. One node machine with 4G
2. Four node machine with 8G
4.2 Result for Test 1
Test 1: counter comparison
This is a test with hacked kernel to record times function
get_any_partial() is invoked and times the inner loop iterates. By
comparing the ratio of two counters, we get to know how many inner
loops we skipped.
Here is a snip of the test patch.
---
static void *get_any_partial() {
get_partial_count++;
do {
for_each_zone_zonelist() {
get_partial_try_count++;
}
} while();
return NULL;
}
---
The result of (get_partial_count / get_partial_try_count):
+----------+----------------+------------+-------------+
| | Base | Patched | Improvement|
+----------+----------------+------------+-------------+
|One Node | 1:3 | 1:0 | - 100% |
+----------+----------------+------------+-------------+
|Four Nodes| 1:5.8 | 1:2.5 | - 56% |
+----------+----------------+------------+-------------+
4.3 Result for Test 2
Test 2: kernel build
Command used:
> time make -j8 bzImage
Each version/system configuration combination has four round kernel
build tests. Take the average result of real to compare.
+----------+----------------+------------+-------------+
| | Base | Patched | Improvement|
+----------+----------------+------------+-------------+
|One Node | 4m41s | 4m32s | - 4.47% |
+----------+----------------+------------+-------------+
|Four Nodes| 4m45s | 4m39s | - 2.92% |
+----------+----------------+------------+-------------+
[akpm@linux-foundation.org: rename variable, tweak comment]
Link: http://lkml.kernel.org/r/20181120033119.30013-1-richard.weiyang@gmail.com
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
---
mm/slub.c | 15 +++++++++++++--
1 file changed, 13 insertions(+), 2 deletions(-)
--- a/mm/slub.c~mm-slub-improve-performance-by-skipping-checked-node-in-get_any_partial
+++ a/mm/slub.c
@@ -1877,7 +1877,7 @@ static void *get_partial_node(struct kme
* Get a page from somewhere. Search in increasing NUMA distances.
*/
static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
- struct kmem_cache_cpu *c)
+ struct kmem_cache_cpu *c, int exclude_nid)
{
#ifdef CONFIG_NUMA
struct zonelist *zonelist;
@@ -1915,6 +1915,9 @@ static void *get_any_partial(struct kmem
for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
struct kmem_cache_node *n;
+ if (exclude_nid == zone_to_nid(zone))
+ continue;
+
n = get_node(s, zone_to_nid(zone));
if (n && cpuset_zone_allowed(zone, flags) &&
@@ -1931,6 +1934,14 @@ static void *get_any_partial(struct kmem
return object;
}
}
+ /*
+ * Failed to get an object from this node, either
+ * because
+ * 1. Failure in the above if check
+ * 2. NULL return from get_partial_node()
+ * So skip this node next time.
+ */
+ exclude_nid = zone_to_nid(zone);
}
} while (read_mems_allowed_retry(cpuset_mems_cookie));
#endif
@@ -1955,7 +1966,7 @@ static void *get_partial(struct kmem_cac
if (object || node != NUMA_NO_NODE)
return object;
- return get_any_partial(s, flags, c);
+ return get_any_partial(s, flags, c, searchnode);
}
#ifdef CONFIG_PREEMPT
_
next prev parent reply other threads:[~2018-12-20 22:41 UTC|newest]
Thread overview: 21+ messages / expand[flat|nested] mbox.gz Atom feed top
2018-11-08 1:12 [PATCH] mm/slub: skip node in case there is no slab to acquire Wei Yang
2018-11-09 20:48 ` Andrew Morton
2018-11-09 23:47 ` Wei Yang
2018-11-13 9:12 ` [PATCH v2] " Wei Yang
2018-11-13 13:17 ` [PATCH] " Michal Hocko
2018-11-13 13:26 ` Wei Yang
2018-11-13 13:34 ` Michal Hocko
2018-11-20 3:31 ` [PATCH v2] mm/slub: improve performance by skipping checked node in get_any_partial() Wei Yang
2018-11-22 3:05 ` Andrew Morton
2018-11-22 9:13 ` Wei Yang
2018-11-22 23:41 ` Wei Yang
2018-11-23 13:39 ` Michal Hocko
2018-11-23 13:49 ` Michal Hocko
2018-11-23 15:27 ` Wei Yang
2018-12-20 22:41 ` Andrew Morton [this message]
2018-12-21 0:25 ` Alexander Duyck
2018-12-21 3:29 ` Wei Yang
2018-12-21 1:37 ` Christopher Lameter
2018-12-21 1:37 ` Christopher Lameter
2018-12-21 3:33 ` Wei Yang
2018-12-24 22:03 ` Wei Yang
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20181220144107.9376344c2be687615ea9aa69@linux-foundation.org \
--to=akpm@linux-foundation.org \
--cc=cl@linux.com \
--cc=iamjoonsoo.kim@lge.com \
--cc=linux-mm@kvack.org \
--cc=mhocko@kernel.org \
--cc=penberg@kernel.org \
--cc=richard.weiyang@gmail.com \
--cc=rientjes@google.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).