ppc: RECLAIM_DISTANCE 10?

Hi,
I have just noticed that ppc has RECLAIM_DISTANCE reduced to 10 set by
56608209d34b (powerpc/numa: Set a smaller value for RECLAIM_DISTANCE to
enable zone reclaim). The commit message suggests that the zone reclaim
is desirable for all NUMA configurations.

History has shown that the zone reclaim is more often harmful than
helpful and leads to performance problems. The default RECLAIM_DISTANCE
for generic case has been increased from 20 to 30 around 3.0
(32e45ff43eaf mm: increase RECLAIM_DISTANCE to 30).

I strongly suspect that the patch is incorrect and it should be
reverted. Before I will send a revert I would like to understand what
led to the patch in the first place. I do not see why would PPC use only
LOCAL_DISTANCE and REMOTE_DISTANCE distances and in fact machines I have
seen use different values.

Anton, could you comment please?
-- 
Michal Hocko
SUSE Labs

On Tue, 18 Feb 2014, Michal Hocko wrote:

Hi,
I have just noticed that ppc has RECLAIM_DISTANCE reduced to 10 set by
56608209d34b (powerpc/numa: Set a smaller value for RECLAIM_DISTANCE to
enable zone reclaim). The commit message suggests that the zone reclaim
is desirable for all NUMA configurations.

History has shown that the zone reclaim is more often harmful than
helpful and leads to performance problems. The default RECLAIM_DISTANCE
for generic case has been increased from 20 to 30 around 3.0
(32e45ff43eaf mm: increase RECLAIM_DISTANCE to 30).

I strongly suspect that the patch is incorrect and it should be
reverted. Before I will send a revert I would like to understand what
led to the patch in the first place. I do not see why would PPC use only
LOCAL_DISTANCE and REMOTE_DISTANCE distances and in fact machines I have
seen use different values.

I strongly suspect that the patch is correct since powerpc node distances 
are different than the architectures you're talking about and get doubled 
for every NUMA domain that the hardware supports.

Hi Michal,

On 18.02.2014 [10:06:58 +0100], Michal Hocko wrote:

Hi,
I have just noticed that ppc has RECLAIM_DISTANCE reduced to 10 set by
56608209d34b (powerpc/numa: Set a smaller value for RECLAIM_DISTANCE to
enable zone reclaim). The commit message suggests that the zone reclaim
is desirable for all NUMA configurations.

History has shown that the zone reclaim is more often harmful than
helpful and leads to performance problems. The default RECLAIM_DISTANCE
for generic case has been increased from 20 to 30 around 3.0
(32e45ff43eaf mm: increase RECLAIM_DISTANCE to 30).

Interesting.

I strongly suspect that the patch is incorrect and it should be
reverted. Before I will send a revert I would like to understand what
led to the patch in the first place. I do not see why would PPC use only
LOCAL_DISTANCE and REMOTE_DISTANCE distances and in fact machines I have
seen use different values.

Anton, could you comment please?

I'll let Anton comment here, but in looking into this issue in working
on CONFIG_HAVE_MEMORYLESS_NODE support, I realized that any LPAR with
memoryless nodes will set zone_reclaim_mode to 1. I think we want to
ignore memoryless nodes when we set up the reclaim mode like the
following? I'll send it as a proper patch if you agree?

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5de4337..4f6ff6f 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -1853,8 +1853,9 @@ static void __paginginit init_zone_allows_reclaim(int nid)
 {
        int i;
 
-       for_each_online_node(i)
-               if (node_distance(nid, i) <= RECLAIM_DISTANCE)
+       for_each_online_node(i) {
+               if (node_distance(nid, i) <= RECLAIM_DISTANCE ||
+                                       local_memory_node(nid) != nid)
                        node_set(i, NODE_DATA(nid)->reclaim_nodes);
                else
                        zone_reclaim_mode = 1;

Note, this won't actually do anything if CONFIG_HAVE_MEMORYLESS_NODES is
not set, but if it is, I think semantically it will indicate that
memoryless nodes *have* to reclaim remotely.

And actually the above won't work, because the callpath is

start_kernel -> setup_arch -> paging_init [-> free_area_init_nodes ->
free_area_init_node -> init_zone_allows_reclaim] which is called before
build_all_zonelists. This is a similar ordering problem as I'm having
with the MEMORYLESS_NODE support, will work on it.

Thanks,
Nish

On 18.02.2014 [15:34:05 -0800], Nishanth Aravamudan wrote:

How about the following?

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5de4337..1a0eced 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -1854,7 +1854,8 @@ static void __paginginit init_zone_allows_reclaim(int nid)
        int i;
 
        for_each_online_node(i)
-               if (node_distance(nid, i) <= RECLAIM_DISTANCE)
+               if (node_distance(nid, i) <= RECLAIM_DISTANCE ||
+                                       !NODE_DATA(nid)->node_present_pages)
                        node_set(i, NODE_DATA(nid)->reclaim_nodes);
                else
                        zone_reclaim_mode = 1;

@@ -4901,13 +4902,13 @@ void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
 
        pgdat->node_id = nid;
        pgdat->node_start_pfn = node_start_pfn;
-       init_zone_allows_reclaim(nid);
 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
        get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
 #endif
        calculate_node_totalpages(pgdat, start_pfn, end_pfn,
                                  zones_size, zholes_size);
 
+       init_zone_allows_reclaim(nid);
        alloc_node_mem_map(pgdat);
 #ifdef CONFIG_FLAT_NODE_MEM_MAP
        printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",

I think it's safe to move init_zone_allows_reclaim, because I don't
think any allocates are occurring here that could cause us to reclaim
anyways, right? Moving it allows us to safely reference
node_present_pages.

Thanks,
Nish

On 18.02.2014 [15:58:00 -0800], Nishanth Aravamudan wrote:

err s/nid/i/ above.

-Nish

On Tue, 18 Feb 2014, Nishanth Aravamudan wrote:

 [ I changed the above from NODE_DATA(nid) -> NODE_DATA(i) as you caught 
   so we're looking at the right code. ]

That can't be right, it would allow reclaiming from a memoryless node.  I 
think what you want is

	for_each_online_node(i) {
		if (!node_present_pages(i))
			continue;
		if (node_distance(nid, i) <= RECLAIM_DISTANCE) {
			node_set(i, NODE_DATA(nid)->reclaim_nodes);
			continue;
		}
		/* Always try to reclaim locally */
		zone_reclaim_mode = 1;
	}

but we really should be able to do for_each_node_state(i, N_MEMORY) here 
and memoryless nodes should already be excluded from that mask.

Yeah, this is fine.

On Tue 18-02-14 14:27:11, David Rientjes wrote:

On Tue, 18 Feb 2014, Michal Hocko wrote:

quoted

Hi,
I have just noticed that ppc has RECLAIM_DISTANCE reduced to 10 set by
56608209d34b (powerpc/numa: Set a smaller value for RECLAIM_DISTANCE to
enable zone reclaim). The commit message suggests that the zone reclaim
is desirable for all NUMA configurations.

History has shown that the zone reclaim is more often harmful than
helpful and leads to performance problems. The default RECLAIM_DISTANCE
for generic case has been increased from 20 to 30 around 3.0
(32e45ff43eaf mm: increase RECLAIM_DISTANCE to 30).

I strongly suspect that the patch is incorrect and it should be
reverted. Before I will send a revert I would like to understand what
led to the patch in the first place. I do not see why would PPC use only
LOCAL_DISTANCE and REMOTE_DISTANCE distances and in fact machines I have
seen use different values.

I strongly suspect that the patch is correct since powerpc node distances 
are different than the architectures you're talking about and get doubled 
for every NUMA domain that the hardware supports.

Even if the units of the distance is different on PPC should every NUMA
machine have zone_reclaim enabled? That doesn't right to me.

-- 
Michal Hocko
SUSE Labs

On Wed, 19 Feb 2014, Michal Hocko wrote:

quoted

I strongly suspect that the patch is correct since powerpc node distances 
are different than the architectures you're talking about and get doubled 
for every NUMA domain that the hardware supports.

Even if the units of the distance is different on PPC should every NUMA
machine have zone_reclaim enabled? That doesn't right to me.

In my experience on powerpc it's very correct, there's typically a 
significant latency in remote access and we don't have the benefit of a 
SLIT that actually defines the locality between proximity domains like we 
do on other architectures.  We have had significant issues with thp, for 
example, being allocated remotely instead of pages locally, much more 
drastic than on our x86 machines, particularly AMD machines.

On Tue 18-02-14 15:34:05, Nishanth Aravamudan wrote:

Hi Michal,

On 18.02.2014 [10:06:58 +0100], Michal Hocko wrote:

quoted

Hi,
I have just noticed that ppc has RECLAIM_DISTANCE reduced to 10 set by
56608209d34b (powerpc/numa: Set a smaller value for RECLAIM_DISTANCE to
enable zone reclaim). The commit message suggests that the zone reclaim
is desirable for all NUMA configurations.

History has shown that the zone reclaim is more often harmful than
helpful and leads to performance problems. The default RECLAIM_DISTANCE
for generic case has been increased from 20 to 30 around 3.0
(32e45ff43eaf mm: increase RECLAIM_DISTANCE to 30).

Interesting.

quoted

I strongly suspect that the patch is incorrect and it should be
reverted. Before I will send a revert I would like to understand what
led to the patch in the first place. I do not see why would PPC use only
LOCAL_DISTANCE and REMOTE_DISTANCE distances and in fact machines I have
seen use different values.

Anton, could you comment please?

I'll let Anton comment here, but in looking into this issue in working
on CONFIG_HAVE_MEMORYLESS_NODE support, I realized that any LPAR with
memoryless nodes will set zone_reclaim_mode to 1. I think we want to
ignore memoryless nodes when we set up the reclaim mode like the
following? I'll send it as a proper patch if you agree?

Funny enough, ppc memoryless node setup is what led me to this code.
We had a setup like this:
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
node 0 size: 0 MB
node 0 free: 0 MB
node 2 cpus:
node 2 size: 7168 MB
node 2 free: 6019 MB
node distances:
node   0   2
0:  10  40
2:  40  10

Which ends up enabling zone_reclaim although there is only a single node
with memory. Not that RECLAIM_DISTANCE would make any difference here as
the distance is even above default RECLAIM_DISTANCE.

I think you just want for_each_node_state(nid, N_MEMORY) and skip all
the memory less nodes, no?
-- 
Michal Hocko
SUSE Labs

On Tue 18-02-14 17:43:38, David Rientjes wrote:

On Tue, 18 Feb 2014, Nishanth Aravamudan wrote:

quoted

How about the following?

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5de4337..1a0eced 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -1854,7 +1854,8 @@ static void __paginginit init_zone_allows_reclaim(int nid)
        int i;
 
        for_each_online_node(i)
-               if (node_distance(nid, i) <= RECLAIM_DISTANCE)
+               if (node_distance(nid, i) <= RECLAIM_DISTANCE ||
+                                       !NODE_DATA(i)->node_present_pages)
                        node_set(i, NODE_DATA(nid)->reclaim_nodes);
                else
                        zone_reclaim_mode = 1;

 [ I changed the above from NODE_DATA(nid) -> NODE_DATA(i) as you caught 
   so we're looking at the right code. ]

That can't be right, it would allow reclaiming from a memoryless node.  I 
think what you want is

	for_each_online_node(i) {
		if (!node_present_pages(i))
			continue;
		if (node_distance(nid, i) <= RECLAIM_DISTANCE) {
			node_set(i, NODE_DATA(nid)->reclaim_nodes);
			continue;
		}
		/* Always try to reclaim locally */
		zone_reclaim_mode = 1;
	}

but we really should be able to do for_each_node_state(i, N_MEMORY) here 
and memoryless nodes should already be excluded from that mask.

Agreed! Actually the code I am currently interested in is based on 3.0
kernel where zone_reclaim_mode is set in build_zonelists which relies on
find_next_best_node which iterates only N_HIGH_MEMORY nodes which should
have non 0 present pages.
[...]
-- 
Michal Hocko
SUSE Labs

On Wed 19-02-14 00:20:21, David Rientjes wrote:

On Wed, 19 Feb 2014, Michal Hocko wrote:

quoted

quoted

I strongly suspect that the patch is correct since powerpc node distances 
are different than the architectures you're talking about and get doubled 
for every NUMA domain that the hardware supports.

Even if the units of the distance is different on PPC should every NUMA
machine have zone_reclaim enabled? That doesn't right to me.

In my experience on powerpc it's very correct, there's typically a 
significant latency in remote access and we don't have the benefit of a 
SLIT that actually defines the locality between proximity domains like we 
do on other architectures. 

Interesting. So is the PPC NUMA basically about local vs. very distant?
Should REMOTE_DISTANCE reflect that as well? Or can we have 
distance < REMOTE_DISTANCE and it would still make sense to have
zone_reclaim enabled?

[...]

Thanks!
-- 
Michal Hocko
SUSE Labs

On 19.02.2014 [08:24:38 -0800], Nishanth Aravamudan wrote:

On 18.02.2014 [17:43:38 -0800], David Rientjes wrote:

quoted

On Tue, 18 Feb 2014, Nishanth Aravamudan wrote:

quoted

How about the following?

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5de4337..1a0eced 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -1854,7 +1854,8 @@ static void __paginginit init_zone_allows_reclaim(int nid)
        int i;
 
        for_each_online_node(i)
-               if (node_distance(nid, i) <= RECLAIM_DISTANCE)
+               if (node_distance(nid, i) <= RECLAIM_DISTANCE ||
+                                       !NODE_DATA(i)->node_present_pages)
                        node_set(i, NODE_DATA(nid)->reclaim_nodes);
                else
                        zone_reclaim_mode = 1;

 [ I changed the above from NODE_DATA(nid) -> NODE_DATA(i) as you caught 
   so we're looking at the right code. ]

That can't be right, it would allow reclaiming from a memoryless node.  I 
think what you want is

Gah, you're right.

quoted

	for_each_online_node(i) {
		if (!node_present_pages(i))
			continue;
		if (node_distance(nid, i) <= RECLAIM_DISTANCE) {
			node_set(i, NODE_DATA(nid)->reclaim_nodes);
			continue;
		}
		/* Always try to reclaim locally */
		zone_reclaim_mode = 1;
	}

but we really should be able to do for_each_node_state(i, N_MEMORY) here 
and memoryless nodes should already be excluded from that mask.

Yep, I found that afterwards, which simplifies the logic. I'll add this
to my series :)

In looking at the code, I am wondering about the following:

init_zone_allows_reclaim() is called for each nid from
free_area_init_node(). Which means that calculate_node_totalpages for
other "later" nids and check_for_memory() [which sets up the N_MEMORY
nodemask] hasn't been called yet.

So, would it make sense to pull up the
                /* Any memory on that node */
                if (pgdat->node_present_pages)
                        node_set_state(nid, N_MEMORY);
                check_for_memory(pgdat, nid);
into free_area_init_node()?

Thanks,
Nish

On 19.02.2014 [09:23:13 +0100], Michal Hocko wrote:

On Tue 18-02-14 15:34:05, Nishanth Aravamudan wrote:

quoted

Hi Michal,

On 18.02.2014 [10:06:58 +0100], Michal Hocko wrote:

quoted

Hi,
I have just noticed that ppc has RECLAIM_DISTANCE reduced to 10 set by
56608209d34b (powerpc/numa: Set a smaller value for RECLAIM_DISTANCE to
enable zone reclaim). The commit message suggests that the zone reclaim
is desirable for all NUMA configurations.

History has shown that the zone reclaim is more often harmful than
helpful and leads to performance problems. The default RECLAIM_DISTANCE
for generic case has been increased from 20 to 30 around 3.0
(32e45ff43eaf mm: increase RECLAIM_DISTANCE to 30).

Interesting.

quoted

I strongly suspect that the patch is incorrect and it should be
reverted. Before I will send a revert I would like to understand what
led to the patch in the first place. I do not see why would PPC use only
LOCAL_DISTANCE and REMOTE_DISTANCE distances and in fact machines I have
seen use different values.

Anton, could you comment please?

I'll let Anton comment here, but in looking into this issue in working
on CONFIG_HAVE_MEMORYLESS_NODE support, I realized that any LPAR with
memoryless nodes will set zone_reclaim_mode to 1. I think we want to
ignore memoryless nodes when we set up the reclaim mode like the
following? I'll send it as a proper patch if you agree?

Funny enough, ppc memoryless node setup is what led me to this code.
We had a setup like this:
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
node 0 size: 0 MB
node 0 free: 0 MB
node 2 cpus:
node 2 size: 7168 MB
node 2 free: 6019 MB
node distances:
node   0   2
0:  10  40
2:  40  10

Yeah, I think this happens fairly often ... and we didn't properly
support it (particularly with SLUB) on powerpc. I'll cc you on my
patchset.

Which ends up enabling zone_reclaim although there is only a single node
with memory. Not that RECLAIM_DISTANCE would make any difference here as
the distance is even above default RECLAIM_DISTANCE.

quoted

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5de4337..4f6ff6f 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -1853,8 +1853,9 @@ static void __paginginit init_zone_allows_reclaim(int nid)
 {
        int i;
 
-       for_each_online_node(i)
-               if (node_distance(nid, i) <= RECLAIM_DISTANCE)
+       for_each_online_node(i) {
+               if (node_distance(nid, i) <= RECLAIM_DISTANCE ||
+                                       local_memory_node(nid) != nid)
                        node_set(i, NODE_DATA(nid)->reclaim_nodes);
                else
                        zone_reclaim_mode = 1;

Note, this won't actually do anything if CONFIG_HAVE_MEMORYLESS_NODES is
not set, but if it is, I think semantically it will indicate that
memoryless nodes *have* to reclaim remotely.

And actually the above won't work, because the callpath is

start_kernel -> setup_arch -> paging_init [-> free_area_init_nodes ->
free_area_init_node -> init_zone_allows_reclaim] which is called before
build_all_zonelists. This is a similar ordering problem as I'm having
with the MEMORYLESS_NODE support, will work on it.

I think you just want for_each_node_state(nid, N_MEMORY) and skip all
the memory less nodes, no?

Yep, thanks!
-Nish

On 18.02.2014 [17:43:38 -0800], David Rientjes wrote:

On Tue, 18 Feb 2014, Nishanth Aravamudan wrote:

quoted

How about the following?

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5de4337..1a0eced 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -1854,7 +1854,8 @@ static void __paginginit init_zone_allows_reclaim(int nid)
        int i;
 
        for_each_online_node(i)
-               if (node_distance(nid, i) <= RECLAIM_DISTANCE)
+               if (node_distance(nid, i) <= RECLAIM_DISTANCE ||
+                                       !NODE_DATA(i)->node_present_pages)
                        node_set(i, NODE_DATA(nid)->reclaim_nodes);
                else
                        zone_reclaim_mode = 1;

 [ I changed the above from NODE_DATA(nid) -> NODE_DATA(i) as you caught 
   so we're looking at the right code. ]

That can't be right, it would allow reclaiming from a memoryless node.  I 
think what you want is

Gah, you're right.

	for_each_online_node(i) {
		if (!node_present_pages(i))
			continue;
		if (node_distance(nid, i) <= RECLAIM_DISTANCE) {
			node_set(i, NODE_DATA(nid)->reclaim_nodes);
			continue;
		}
		/* Always try to reclaim locally */
		zone_reclaim_mode = 1;
	}

but we really should be able to do for_each_node_state(i, N_MEMORY) here 
and memoryless nodes should already be excluded from that mask.

Yep, I found that afterwards, which simplifies the logic. I'll add this
to my series :)

<snip>

quoted

I think it's safe to move init_zone_allows_reclaim, because I don't
think any allocates are occurring here that could cause us to reclaim
anyways, right? Moving it allows us to safely reference
node_present_pages.

Yeah, this is fine.

Thanks,
Nish

On Wed, 19 Feb 2014, Michal Hocko wrote:

Interesting. So is the PPC NUMA basically about local vs. very distant?

The point is that it's impossible to tell how distant they are from one 
NUMA domain to the next NUMA domain.

Should REMOTE_DISTANCE reflect that as well? Or can we have 
distance < REMOTE_DISTANCE and it would still make sense to have
zone_reclaim enabled?

Ppc doesn't want to allocate in a different NUMA domain unless required, 
the latency of a remote access is too high.  Everything that isn't in the 
same domain has a distance >10 and is setup as 2^(domain hops) * 10.  We 
don't have the ability like with a SLIT to define remote nodes to have 
local latency vs remote or costly latency so the safe setting is a 
RECLAIM_DISTANCE of 10.

On Wed 19-02-14 08:33:45, Nishanth Aravamudan wrote:

On 19.02.2014 [08:24:38 -0800], Nishanth Aravamudan wrote:

quoted

On 18.02.2014 [17:43:38 -0800], David Rientjes wrote:

quoted

On Tue, 18 Feb 2014, Nishanth Aravamudan wrote:

quoted

How about the following?

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5de4337..1a0eced 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -1854,7 +1854,8 @@ static void __paginginit init_zone_allows_reclaim(int nid)
        int i;
 
        for_each_online_node(i)
-               if (node_distance(nid, i) <= RECLAIM_DISTANCE)
+               if (node_distance(nid, i) <= RECLAIM_DISTANCE ||
+                                       !NODE_DATA(i)->node_present_pages)
                        node_set(i, NODE_DATA(nid)->reclaim_nodes);
                else
                        zone_reclaim_mode = 1;

 [ I changed the above from NODE_DATA(nid) -> NODE_DATA(i) as you caught 
   so we're looking at the right code. ]

That can't be right, it would allow reclaiming from a memoryless node.  I 
think what you want is

Gah, you're right.

quoted

	for_each_online_node(i) {
		if (!node_present_pages(i))
			continue;
		if (node_distance(nid, i) <= RECLAIM_DISTANCE) {
			node_set(i, NODE_DATA(nid)->reclaim_nodes);
			continue;
		}
		/* Always try to reclaim locally */
		zone_reclaim_mode = 1;
	}

but we really should be able to do for_each_node_state(i, N_MEMORY) here 
and memoryless nodes should already be excluded from that mask.

Yep, I found that afterwards, which simplifies the logic. I'll add this
to my series :)

In looking at the code, I am wondering about the following:

init_zone_allows_reclaim() is called for each nid from
free_area_init_node(). Which means that calculate_node_totalpages for
other "later" nids and check_for_memory() [which sets up the N_MEMORY
nodemask] hasn't been called yet.

So, would it make sense to pull up the
                /* Any memory on that node */
                if (pgdat->node_present_pages)
                        node_set_state(nid, N_MEMORY);
                check_for_memory(pgdat, nid);
into free_area_init_node()?

Dunno, but it shouldn't be needed because nodes are set N_MEMORY earlier
in early_calculate_totalpages as mentioned in other email.

-- 
Michal Hocko
SUSE Labs