[PATCH -V7 0/3] numa balancing: Migrate on fault among multiple bound nodes

[PATCH -V7 0/3] numa balancing: Migrate on fault among multiple bound nodes

[Huang Ying]

To make it possible to optimize cross-socket memory accessing with
AutoNUMA even if the memory of the application is bound to multiple
NUMA nodes.

Patch [2/3] and [3/3] are NOT kernel patches.  Instead, they are
patches for man-pages and numactl respectively.  They are sent
together to make it easy to review the newly added kernel API.

Changes:

v7:

- Make set_mempolicy() return -1 with errno is set to EINVAL if mode
  isn't MPOL_BIND per Mel's comments.  Revise document accordingly
  too.

v6:

- Rebased on latest upstream kernel 5.10-rc5

- Added some benchmark data and example in patch description of [1/3]

- Rename AutoNUMA to NUMA Balancing

- Add patches to man-pages [2/3] and numactl [3/3]

v5:

- Remove mbind() support, because it's not clear that it's necessary.

v4:

- Use new flags instead of reuse MPOL_MF_LAZY.

v3:

- Rebased on latest upstream (v5.10-rc3)

- Revised the change log.

v2:

- Rebased on latest upstream (v5.10-rc1)

Best Regards,
Huang, Ying

[PATCH -V7 1/3] numa balancing: Migrate on fault among multiple bound nodes

[Huang Ying]

Now, NUMA balancing can only optimize the page placement among the
NUMA nodes if the default memory policy is used.  Because the memory
policy specified explicitly should take precedence.  But this seems
too strict in some situations.  For example, on a system with 4 NUMA
nodes, if the memory of an application is bound to the node 0 and 1,
NUMA balancing can potentially migrate the pages between the node 0
and 1 to reduce cross-node accessing without breaking the explicit
memory binding policy.

So in this patch, we add MPOL_F_NUMA_BALANCING mode flag to
set_mempolicy() when mode is MPOL_BIND.  With the flag specified, NUMA
balancing will be enabled within the thread to optimize the page
placement within the constrains of the specified memory binding
policy.  With the newly added flag, the NUMA balancing control
mechanism becomes,

- sysctl knob numa_balancing can enable/disable the NUMA balancing
  globally.

- even if sysctl numa_balancing is enabled, the NUMA balancing will be
  disabled for the memory areas or applications with the explicit memory
  policy by default.

- MPOL_F_NUMA_BALANCING can be used to enable the NUMA balancing for the
  applications when specifying the explicit memory policy (MPOL_BIND).

Various page placement optimization based on the NUMA balancing can be
done with these flags.  As the first step, in this patch, if the
memory of the application is bound to multiple nodes (MPOL_BIND), and
in the hint page fault handler the accessing node are in the policy
nodemask, the page will be tried to be migrated to the accessing node
to reduce the cross-node accessing.

If the newly added MPOL_F_NUMA_BALANCING flag is specified by an
application on an old kernel version without its support,
set_mempolicy() will return -1 and errno will be set to EINVAL.  The
application can use this behavior to run on both old and new kernel
versions.

And if the MPOL_F_NUMA_BALANCING flag is specified for the mode other
than MPOL_BIND, set_mempolicy() will return -1 and errno will be set
to EINVAL as before.  Because we don't support optimization based on
the NUMA balancing for these modes.

In the previous version of the patch, we tried to reuse MPOL_MF_LAZY
for mbind().  But that flag is tied to MPOL_MF_MOVE.*, so it seems not
a good API/ABI for the purpose of the patch.

And because it's not clear whether it's necessary to enable NUMA
balancing for a specific memory area inside an application, so we only
add the flag at the thread level (set_mempolicy()) instead of the
memory area level (mbind()).  We can do that when it become necessary.

To test the patch, we run a test case as follows on a 4-node machine
with 192 GB memory (48 GB per node).

1. Change pmbench memory accessing benchmark to call set_mempolicy()
   to bind its memory to node 1 and 3 and enable NUMA balancing.  Some
   related code snippets are as follows,

     #include <numaif.h>
     #include <numa.h>

	struct bitmask *bmp;
	int ret;

	bmp = numa_parse_nodestring("1,3");
	ret = set_mempolicy(MPOL_BIND | MPOL_F_NUMA_BALANCING,
			    bmp->maskp, bmp->size + 1);
	/* If MPOL_F_NUMA_BALANCING isn't supported, fall back to MPOL_BIND */
	if (ret < 0 && errno == EINVAL)
		ret = set_mempolicy(MPOL_BIND, bmp->maskp, bmp->size + 1);
	if (ret < 0) {
		perror("Failed to call set_mempolicy");
		exit(-1);
	}

2. Run a memory eater on node 3 to use 40 GB memory before running pmbench.

3. Run pmbench with 64 processes, the working-set size of each process
   is 640 MB, so the total working-set size is 64 * 640 MB = 40 GB.  The
   CPU and the memory (as in step 1.) of all pmbench processes is bound
   to node 1 and 3. So, after CPU usage is balanced, some pmbench
   processes run on the CPUs of the node 3 will access the memory of
   the node 1.

4. After the pmbench processes run for 100 seconds, kill the memory
   eater.  Now it's possible for some pmbench processes to migrate
   their pages from node 1 to node 3 to reduce cross-node accessing.

Test results show that, with the patch, the pages can be migrated from
node 1 to node 3 after killing the memory eater, and the pmbench score
can increase about 17.5%.

Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: David Rientjes <rientjes@google.com>
Cc: linux-api@vger.kernel.org
---
 include/uapi/linux/mempolicy.h |  4 +++-
 mm/mempolicy.c                 | 16 ++++++++++++++++
 2 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/include/uapi/linux/mempolicy.h b/include/uapi/linux/mempolicy.h
index 3354774af61e..8948467b3992 100644
--- a/include/uapi/linux/mempolicy.h
+++ b/include/uapi/linux/mempolicy.h
@@ -28,12 +28,14 @@ enum {
 /* Flags for set_mempolicy */
 #define MPOL_F_STATIC_NODES	(1 << 15)
 #define MPOL_F_RELATIVE_NODES	(1 << 14)
+#define MPOL_F_NUMA_BALANCING	(1 << 13) /* Optimize with NUMA balancing if possible */
 
 /*
  * MPOL_MODE_FLAGS is the union of all possible optional mode flags passed to
  * either set_mempolicy() or mbind().
  */
-#define MPOL_MODE_FLAGS	(MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES)
+#define MPOL_MODE_FLAGS							\
+	(MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES | MPOL_F_NUMA_BALANCING)
 
 /* Flags for get_mempolicy */
 #define MPOL_F_NODE	(1<<0)	/* return next IL mode instead of node mask */
diff --git a/mm/mempolicy.c b/mm/mempolicy.c
index 3ca4898f3f24..c3f70d8e17d6 100644
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -875,6 +875,16 @@ static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 		goto out;
 	}
 
+	if (flags & MPOL_F_NUMA_BALANCING) {
+		if (new && new->mode == MPOL_BIND) {
+			new->flags |= (MPOL_F_MOF | MPOL_F_MORON);
+		} else {
+			ret = -EINVAL;
+			mpol_put(new);
+			goto out;
+		}
+	}
+
 	ret = mpol_set_nodemask(new, nodes, scratch);
 	if (ret) {
 		mpol_put(new);
@@ -2490,6 +2500,12 @@ int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long
 		break;
 
 	case MPOL_BIND:
+		/* Optimize placement among multiple nodes via NUMA balancing */
+		if (pol->flags & MPOL_F_MORON) {
+			if (node_isset(thisnid, pol->v.nodes))
+				break;
+			goto out;
+		}
 
 		/*
 		 * allows binding to multiple nodes.
-- 
2.29.2

[PATCH -V7 2/3] NOT kernel/man2/set_mempolicy.2: Add mode flag MPOL_F_NUMA_BALANCING

[Huang Ying]

Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
---
 man2/set_mempolicy.2 | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/man2/set_mempolicy.2 b/man2/set_mempolicy.2
index 68011eecb..fb2e6fd96 100644
--- a/man2/set_mempolicy.2
+++ b/man2/set_mempolicy.2
@@ -113,6 +113,15 @@ A nonempty
 .I nodemask
 specifies node IDs that are relative to the set of
 node IDs allowed by the process's current cpuset.
+.TP
+.BR MPOL_F_NUMA_BALANCING " (since Linux 5.11)"
+When
+.I mode
+is MPOL_BIND, enable the Linux kernel NUMA balancing for the task if
+it is supported by kernel.
+If the flag isn't supported by Linux kernel, or is used with
+.I mode
+other than MPOL_BIND, return -1 and errno is set to EINVAL.
 .PP
 .I nodemask
 points to a bit mask of node IDs that contains up to
@@ -293,6 +302,11 @@ argument specified both
 .B MPOL_F_STATIC_NODES
 and
 .BR MPOL_F_RELATIVE_NODES .
+Or, the
+.B MPOL_F_NUMA_BALANCING
+isn't supported by the Linux kernel, or is used with
+.I mode
+other than MPOL_BIND.
 .TP
 .B ENOMEM
 Insufficient kernel memory was available.
-- 
2.29.2

[PATCH -V7 3/3] NOT kernel/numactl: Support to enable Linux kernel NUMA balancing

[Huang Ying]

A new API: numa_set_membind_balancing() is added to libnuma.  It is
same as numa_set_membind() except that the Linux kernel NUMA balancing
will be enabled for the task if the feature is supported by the
kernel.

At the same time, a new option: --balancing (-b) is added to numactl.
Which can be used before the --membind/-m memory policy in the command
line.  With it, the Linux kernel NUMA balancing will be enabled for
the process if --membind/-m is used and the feature is supported by
the kernel.

Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
---
 libnuma.c         | 14 ++++++++++++++
 numa.3            | 15 +++++++++++++++
 numa.h            |  4 ++++
 numactl.8         | 12 ++++++++++++
 numactl.c         | 17 ++++++++++++++---
 numaif.h          |  3 +++
 versions.ldscript |  8 ++++++++
 7 files changed, 70 insertions(+), 3 deletions(-)

diff --git a/libnuma.c b/libnuma.c
index 88f479b..f073c50 100644
--- a/libnuma.c
+++ b/libnuma.c
@@ -1064,6 +1064,20 @@ numa_set_membind_v2(struct bitmask *bmp)
 
 make_internal_alias(numa_set_membind_v2);
 
+void
+numa_set_membind_balancing(struct bitmask *bmp)
+{
+	/* MPOL_F_NUMA_BALANCING: ignore if unsupported */
+	if (set_mempolicy(MPOL_BIND | MPOL_F_NUMA_BALANCING,
+			  bmp->maskp, bmp->size + 1) < 0) {
+		if (errno == EINVAL) {
+			errno = 0;
+			numa_set_membind_v2(bmp);
+		} else
+			numa_error("set_mempolicy");
+	}
+}
+
 /*
  * copy a bitmask map body to a numa.h nodemask_t structure
  */
diff --git a/numa.3 b/numa.3
index 3e18098..af01c8f 100644
--- a/numa.3
+++ b/numa.3
@@ -80,6 +80,8 @@ numa \- NUMA policy library
 .br
 .BI "void numa_set_membind(struct bitmask *" nodemask );
 .br
+.BI "void numa_set_membind_balancing(struct bitmask *" nodemask );
+.br
 .B struct bitmask *numa_get_membind(void);
 .sp
 .BI "void *numa_alloc_onnode(size_t " size ", int " node );
@@ -538,6 +540,19 @@ that contains nodes other than those in the mask returned by
 .IR numa_get_mems_allowed ()
 will result in an error.
 
+.BR numa_set_membind_balancing ()
+sets the memory allocation mask and enable the Linux kernel NUMA
+balancing for the task if the feature is supported by the kernel.
+The task will only allocate memory from the nodes set in
+.IR nodemask .
+Passing an empty
+.I nodemask
+or a
+.I nodemask
+that contains nodes other than those in the mask returned by
+.IR numa_get_mems_allowed ()
+will result in an error.
+
 .BR numa_get_membind ()
 returns the mask of nodes from which memory can currently be allocated.
 If the returned mask is equal to
diff --git a/numa.h b/numa.h
index bd1d676..5d8543a 100644
--- a/numa.h
+++ b/numa.h
@@ -192,6 +192,10 @@ void numa_set_localalloc(void);
 /* Only allocate memory from the nodes set in mask. 0 to turn off */
 void numa_set_membind(struct bitmask *nodemask);
 
+/* Only allocate memory from the nodes set in mask. Optimize page
+   placement with Linux kernel NUMA balancing if possible. 0 to turn off */
+void numa_set_membind_balancing(struct bitmask *bmp);
+
 /* Return current membind */
 struct bitmask *numa_get_membind(void);
 
diff --git a/numactl.8 b/numactl.8
index f3bb22b..7d52688 100644
--- a/numactl.8
+++ b/numactl.8
@@ -25,6 +25,8 @@ numactl \- Control NUMA policy for processes or shared memory
 [
 .B \-\-all
 ] [
+.B \-\-balancing
+] [
 .B \-\-interleave nodes
 ] [
 .B \-\-preferred node 
@@ -168,6 +170,12 @@ but if memory cannot be allocated there fall back to other nodes.
 This option takes only a single node number.
 Relative notation may be used.
 .TP
+.B \-\-balancing, \-b
+Enable Linux kernel NUMA balancing for the process if it is supported by kernel.
+This should only be used with
+.I \-\-membind, \-m
+only, otherwise ignored.
+.TP
 .B \-\-show, \-s
 Show NUMA policy settings of the current process. 
 .TP
@@ -278,6 +286,10 @@ numactl \-\-cpunodebind=0 \-\-membind=0,1 -- process -l
 Run process as above, but with an option (-l) that would be confused with
 a numactl option.
 
+numactl \-\-cpunodebind=0 \-\-balancing \-\-membind=0,1 process
+Run process on node 0 with memory allocated on node 0 and 1.  Optimize the
+page placement with Linux kernel NUMA balancing mechanism if possible.
+
 numactl \-\-cpunodebind=netdev:eth0 \-\-membind=netdev:eth0 network-server
 Run network-server on the node of network device eth0 with its memory
 also in the same node.
diff --git a/numactl.c b/numactl.c
index df9dbcb..5a9d2df 100644
--- a/numactl.c
+++ b/numactl.c
@@ -45,6 +45,7 @@ struct option opts[] = {
 	{"membind", 1, 0, 'm'},
 	{"show", 0, 0, 's' },
 	{"localalloc", 0,0, 'l'},
+	{"balancing", 0, 0, 'b'},
 	{"hardware", 0,0,'H' },
 
 	{"shm", 1, 0, 'S'},
@@ -65,9 +66,10 @@ struct option opts[] = {
 void usage(void)
 {
 	fprintf(stderr,
-		"usage: numactl [--all | -a] [--interleave= | -i <nodes>] [--preferred= | -p <node>]\n"
-		"               [--physcpubind= | -C <cpus>] [--cpunodebind= | -N <nodes>]\n"
-		"               [--membind= | -m <nodes>] [--localalloc | -l] command args ...\n"
+		"usage: numactl [--all | -a] [--balancing | -b] [--interleave= | -i <nodes>]\n"
+		"               [--preferred= | -p <node>] [--physcpubind= | -C <cpus>]\n"
+		"               [--cpunodebind= | -N <nodes>] [--membind= | -m <nodes>]\n"
+		"               [--localalloc | -l] command args ...\n"
 		"       numactl [--show | -s]\n"
 		"       numactl [--hardware | -H]\n"
 		"       numactl [--length | -l <length>] [--offset | -o <offset>] [--shmmode | -M <shmmode>]\n"
@@ -90,6 +92,8 @@ void usage(void)
 		"all numbers and ranges can be made cpuset-relative with +\n"
 		"the old --cpubind argument is deprecated.\n"
 		"use --cpunodebind or --physcpubind instead\n"
+		"use --balancing | -b to enable Linux kernel NUMA balancing\n"
+		"for the process if it is supported by kernel\n"
 		"<length> can have g (GB), m (MB) or k (KB) suffixes\n");
 	exit(1);
 }
@@ -338,6 +342,7 @@ int do_dump = 0;
 int shmattached = 0;
 int did_node_cpu_parse = 0;
 int parse_all = 0;
+int numa_balancing = 0;
 char *shmoption;
 
 void check_cpubind(int flag)
@@ -431,6 +436,10 @@ int main(int ac, char **av)
 			nopolicy();
 			hardware();
 			exit(0);
+		case 'b': /* --balancing  */
+			nopolicy();
+			numa_balancing = 1;
+			break;
 		case 'i': /* --interleave */
 			checknuma();
 			if (parse_all)
@@ -507,6 +516,8 @@ int main(int ac, char **av)
 			numa_set_bind_policy(1);
 			if (shmfd >= 0) {
 				numa_tonodemask_memory(shmptr, shmlen, mask);
+			} else if (numa_balancing) {
+				numa_set_membind_balancing(mask);
 			} else {
 				numa_set_membind(mask);
 			}
diff --git a/numaif.h b/numaif.h
index 91aa230..32c12c3 100644
--- a/numaif.h
+++ b/numaif.h
@@ -29,6 +29,9 @@ extern long move_pages(int pid, unsigned long count,
 #define MPOL_LOCAL       4
 #define MPOL_MAX         5
 
+/* Flags for set_mempolicy, specified in mode */
+#define MPOL_F_NUMA_BALANCING	(1 << 13) /* Optimize with NUMA balancing if possible */
+
 /* Flags for get_mem_policy */
 #define MPOL_F_NODE    (1<<0)   /* return next il node or node of address */
 				/* Warning: MPOL_F_NODE is unsupported and
diff --git a/versions.ldscript b/versions.ldscript
index 23074a0..358eeeb 100644
--- a/versions.ldscript
+++ b/versions.ldscript
@@ -146,3 +146,11 @@ libnuma_1.4 {
   local:
     *;
 } libnuma_1.3;
+
+# New interface for membind with NUMA balancing optimization
+libnuma_1.5 {
+  global:
+    numa_set_membind_balancing;
+  local:
+    *;
+} libnuma_1.4;
-- 
2.29.2

Re: [PATCH -V7 2/3] NOT kernel/man2/set_mempolicy.2: Add mode flag MPOL_F_NUMA_BALANCING

[Alejandro Colomar (man-pages)]

Hi Huang Ying,

Please, see a few fixes below.

Thanks,

Alex

On 12/4/20 10:15 AM, Huang Ying wrote:
> Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
> ---
>  man2/set_mempolicy.2 | 14 ++++++++++++++
>  1 file changed, 14 insertions(+)
> 
> diff --git a/man2/set_mempolicy.2 b/man2/set_mempolicy.2
> index 68011eecb..fb2e6fd96 100644
> --- a/man2/set_mempolicy.2
> +++ b/man2/set_mempolicy.2
> @@ -113,6 +113,15 @@ A nonempty
>  .I nodemask
>  specifies node IDs that are relative to the set of
>  node IDs allowed by the process's current cpuset.
> +.TP
> +.BR MPOL_F_NUMA_BALANCING " (since Linux 5.11)"
> +When
> +.I mode
> +is MPOL_BIND, enable the Linux kernel NUMA balancing for the task if

.B MPOL_BIND

> +it is supported by kernel.
> +If the flag isn't supported by Linux kernel, or is used with
> +.I mode> +other than MPOL_BIND, return -1 and errno is set to EINVAL.

.BR MPOL_BIND ,

A minus sign should be escaped:
\-1
See man-pages(7)::STYLE GUIDE::Generating optimal glyphs)

.I errno
.BR EINVAL .

>  .PP
>  .I nodemask
>  points to a bit mask of node IDs that contains up to
> @@ -293,6 +302,11 @@ argument specified both
>  .B MPOL_F_STATIC_NODES
>  and
>  .BR MPOL_F_RELATIVE_NODES .
> +Or, the
> +.B MPOL_F_NUMA_BALANCING
> +isn't supported by the Linux kernel, or is used with
> +.I mode
> +other than MPOL_BIND.

.BR MPOL_BIND .

>  .TP
>  .B ENOMEM
>  Insufficient kernel memory was available.
> 

Re: [PATCH -V7 2/3] NOT kernel/man2/set_mempolicy.2: Add mode flag MPOL_F_NUMA_BALANCING

[Huang, Ying]

Hi, Alex,

"Alejandro Colomar (man-pages)" <alx.manpages@gmail.com> writes:

> Hi Huang Ying,
>
> Please, see a few fixes below.
>
> Thanks,
>
> Alex
>
> On 12/4/20 10:15 AM, Huang Ying wrote:
>> Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
>> ---
>>  man2/set_mempolicy.2 | 14 ++++++++++++++
>>  1 file changed, 14 insertions(+)
>> 
>> diff --git a/man2/set_mempolicy.2 b/man2/set_mempolicy.2
>> index 68011eecb..fb2e6fd96 100644
>> --- a/man2/set_mempolicy.2
>> +++ b/man2/set_mempolicy.2
>> @@ -113,6 +113,15 @@ A nonempty
>>  .I nodemask
>>  specifies node IDs that are relative to the set of
>>  node IDs allowed by the process's current cpuset.
>> +.TP
>> +.BR MPOL_F_NUMA_BALANCING " (since Linux 5.11)"
>> +When
>> +.I mode
>> +is MPOL_BIND, enable the Linux kernel NUMA balancing for the task if
>
> .B MPOL_BIND
>
>> +it is supported by kernel.
>> +If the flag isn't supported by Linux kernel, or is used with
>> +.I mode> +other than MPOL_BIND, return -1 and errno is set to EINVAL.
>
> .BR MPOL_BIND ,
>
> A minus sign should be escaped:
> \-1
> See man-pages(7)::STYLE GUIDE::Generating optimal glyphs)
>
> .I errno
> .BR EINVAL .
>
>>  .PP
>>  .I nodemask
>>  points to a bit mask of node IDs that contains up to
>> @@ -293,6 +302,11 @@ argument specified both
>>  .B MPOL_F_STATIC_NODES
>>  and
>>  .BR MPOL_F_RELATIVE_NODES .
>> +Or, the
>> +.B MPOL_F_NUMA_BALANCING
>> +isn't supported by the Linux kernel, or is used with
>> +.I mode
>> +other than MPOL_BIND.
>
> .BR MPOL_BIND .
>
>>  .TP
>>  .B ENOMEM
>>  Insufficient kernel memory was available.
>> 

Thanks a lot for your comments!  Will change them in the next version.

Best Regards,
Huang, Ying