Skip to contentSkip to conversation

Fix NUMA nodes fallback list ordering

PATCH Applied Bharata B Rao 4 participants 13 messages Aug 30, 2021 – Sep 3, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover Fix NUMA nodes fallback list ordering

For a NUMA system that has multiple nodes at same distance from other nodes, the fallback list generation prefers same node order for them instead of round-robin thereby penalizing one node over others. This series fixes it.

More description of the problem and the fix is present in the patch description.

Bharata B Rao (1):
mm/page_alloc: Print node fallback order

Krupa Ramakrishnan (1):
mm/page_alloc: Use accumulated load when building node fallback list

 mm/page_alloc.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH 1/2 mm/page_alloc: Print node fallback order

Print information message about the allocation fallback order for each NUMA node during boot.

No functional changes here. This makes it easier to illustrate the problem in the node fallback list generation, which the next patch fixes.

---
 mm/page_alloc.c | 4 ++++
 1 file changed, 4 insertions(+)
mm/page_alloc.c +4 -0
index eeb3a9cb36bb..22f7ad6ec11c 100644
@@ -6277,6 +6277,10 @@ static void build_zonelists(pg_data_t *pgdat)
 
 	build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
 	build_thisnode_zonelists(pgdat);
+	pr_info("Fallback order for Node %d: ", local_node);
+	for (node = 0; node < nr_nodes; node++)
+		pr_cont("%d ", node_order[node]);
+	pr_cont("\n");
 }
 
 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
Signature · 1 line
2.25.1
Acked-by on v1 1/2
#2 lore raw

Re: [FIX PATCH 1/2] mm/page_alloc: Print node fallback order

Quoted 8 lines · Bharata B Rao

Print information message about the allocation fallback order for each NUMA node during boot.

No functional changes here. This makes it easier to illustrate the problem in the node fallback list generation, which the next patch fixes.

While the message is just informational, I think it's possible that the bug could have been found earlier if this message was present so...

Signature · 2 lines
Mel Gorman
SUSE Labs
#3 lore raw

Re: [FIX PATCH 1/2] mm/page_alloc: Print node fallback order

In reply to Bharata B Rao
Quoted 18 lines · Bharata B Rao

Print information message about the allocation fallback order for each NUMA node during boot.

No functional changes here. This makes it easier to illustrate the problem in the node fallback list generation, which the next patch fixes.

---
 mm/page_alloc.c | 4 ++++
 1 file changed, 4 insertions(+)
mm/page_alloc.c +0 -0
index eeb3a9cb36bb..22f7ad6ec11c 100644

mm/page_alloc.c static void build_zonelists(pg_data_t *pgdat) · line 6286

Quoted 10 lines
mm/page_alloc.c +4 -0
@@ -6277,6 +6277,10 @@ static void build_zonelists(pg_data_t *pgdat)
 
 	build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
 	build_thisnode_zonelists(pgdat);
+	pr_info("Fallback order for Node %d: ", local_node);
+	for (node = 0; node < nr_nodes; node++)
+		pr_cont("%d ", node_order[node]);
+	pr_cont("\n");
 }
 
 #ifdef CONFIG_HAVE_MEMORYLESS_NODES

A small nit, checkpatch.pl throws up an warning. Should this use pr_info() instead ?

WARNING: Avoid logging continuation uses where feasible #29: FILE: mm/page_alloc.c:6282:
+ pr_cont("%d ", node_order[node]);

WARNING: Avoid logging continuation uses where feasible #30: FILE: mm/page_alloc.c:6283:
+ pr_cont("\n");

total: 0 errors, 2 warnings, 10 lines checked

#4 lore raw

Re: [FIX PATCH 1/2] mm/page_alloc: Print node fallback order

Quoted 39 lines · Anshuman Khandual
Quoted 18 lines · Bharata B Rao

Print information message about the allocation fallback order for each NUMA node during boot.

No functional changes here. This makes it easier to illustrate the problem in the node fallback list generation, which the next patch fixes.

---
 mm/page_alloc.c | 4 ++++
 1 file changed, 4 insertions(+)
mm/page_alloc.c +0 -0
index eeb3a9cb36bb..22f7ad6ec11c 100644

mm/page_alloc.c static void build_zonelists(pg_data_t *pgdat) · line 6286

Quoted 10 lines
mm/page_alloc.c +4 -0
@@ -6277,6 +6277,10 @@ static void build_zonelists(pg_data_t *pgdat)
 
 	build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
 	build_thisnode_zonelists(pgdat);
+	pr_info("Fallback order for Node %d: ", local_node);
+	for (node = 0; node < nr_nodes; node++)
+		pr_cont("%d ", node_order[node]);
+	pr_cont("\n");
 }
 
 #ifdef CONFIG_HAVE_MEMORYLESS_NODES

A small nit, checkpatch.pl throws up an warning. Should this use pr_info() instead ?

WARNING: Avoid logging continuation uses where feasible #29: FILE: mm/page_alloc.c:6282:
+ pr_cont("%d ", node_order[node]);

Quoted 5 lines

WARNING: Avoid logging continuation uses where feasible #30: FILE: mm/page_alloc.c:6283:
+ pr_cont("\n");

total: 0 errors, 2 warnings, 10 lines checked

Yes, I am aware of this, but then it made sense for the fallback list to be printed in one line continuously. Hence used pr_cont().

Regards,
Bharata.

Quoted 1 line
Reviewed-by on v1 1/2
#5 lore raw

Re: [FIX PATCH 1/2] mm/page_alloc: Print node fallback order

In reply to Bharata B Rao
Quoted 18 lines · Bharata B Rao

Print information message about the allocation fallback order for each NUMA node during boot.

No functional changes here. This makes it easier to illustrate the problem in the node fallback list generation, which the next patch fixes.

---
 mm/page_alloc.c | 4 ++++
 1 file changed, 4 insertions(+)
mm/page_alloc.c +0 -0
index eeb3a9cb36bb..22f7ad6ec11c 100644

mm/page_alloc.c static void build_zonelists(pg_data_t *pgdat) · line 6286

Quoted 10 lines
mm/page_alloc.c +4 -0
@@ -6277,6 +6277,10 @@ static void build_zonelists(pg_data_t *pgdat)
 
 	build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
 	build_thisnode_zonelists(pgdat);
+	pr_info("Fallback order for Node %d: ", local_node);
+	for (node = 0; node < nr_nodes; node++)
+		pr_cont("%d ", node_order[node]);
+	pr_cont("\n");
 }
 
 #ifdef CONFIG_HAVE_MEMORYLESS_NODES

Node allocation preference sequence for a given accessing node is an important information on large systems. This information during boot and (hotplug --> online sequence) will help the user.

#6 lore raw

PATCH 2/2 mm/page_alloc: Use accumulated load when building node fallback list

In reply to Bharata B Rao

From: Krupa Ramakrishnan <[email protected]>

In build_zonelists(), when the fallback list is built for the nodes, the node load gets reinitialized during each iteration. This results in nodes with same distances occupying the same slot in different node fallback lists rather than appearing in the intended round- robin manner. This results in one node getting picked for allocation more compared to other nodes with the same distance.

As an example, consider a 4 node system with the following distance matrix.

Node 0 1 2 3
----------------

0    10 12 32 32
1    12 10 32 32
2    32 32 10 12
3    32 32 12 10

For this case, the node fallback list gets built like this:

Node Fallback list
---------------------
0 0 1 2 3
1 1 0 3 2
2 2 3 0 1
3 3 2 0 1 <-- Unexpected fallback order

In the fallback list for nodes 2 and 3, the nodes 0 and 1 appear in the same order which results in more allocations getting satisfied from node 0 compared to node 1.

The effect of this on remote memory bandwidth as seen by stream benchmark is shown below:

Case 1: Bandwidth from cores on nodes 2 & 3 to memory on nodes 0 & 1

	(numactl -m 0,1 ./stream_lowOverhead ... --cores <from 2, 3>)
Case 2: Bandwidth from cores on nodes 0 & 1 to memory on nodes 2 & 3
	(numactl -m 2,3 ./stream_lowOverhead ... --cores <from 0, 1>)

----------------------------------------

		BANDWIDTH (MB/s)
    TEST	Case 1		Case 2
----------------------------------------
    COPY	57479.6		110791.8
   SCALE	55372.9		105685.9
     ADD	50460.6		96734.2

TRIADD 50397.6 97119.1
----------------------------------------

The bandwidth drop in Case 1 occurs because most of the allocations get satisfied by node 0 as it appears first in the fallback order for both nodes 2 and 3.

This can be fixed by accumulating the node load in build_zonelists() rather than reinitializing it during each iteration. With this the nodes with the same distance rightly get assigned in the round robin manner. In fact this was how it was originally until the commit f0c0b2b808f2 ("change zonelist order: zonelist order selection logic") dropped the load accumulation and resorted to initializing the load during each iteration. While zonelist ordering was removed by commit c9bff3eebc09 ("mm, page_alloc: rip out ZONELIST_ORDER_ZONE"), the change to the node load accumulation in build_zonelists() remained. So essentially this patch reverts back to the accumulated node load logic.

After this fix, the fallback order gets built like this:

Node Fallback list
------------------

0    0 1 2 3
1    1 0 3 2
2    2 3 0 1
3    3 2 1 0 <-- Note the change here

The bandwidth in Case 1 improves and matches Case 2 as shown below.

----------------------------------------

		BANDWIDTH (MB/s)
    TEST	Case 1		Case 2
----------------------------------------
    COPY	110438.9	110107.2
   SCALE	105930.5	105817.5
     ADD	97005.1		96159.8

TRIADD 97441.5 96757.1
----------------------------------------

The correctness of the fallback list generation has been verified for the above node configuration where the node 3 starts as memory-less node and comes up online only during memory hotplug.

[[email protected]: Added changelog, review, test validation]

Fixes f0c0b2b808f2 ("change zonelist order: zonelist order selection

logic")

---
 mm/page_alloc.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/page_alloc.c +1 -1
index 22f7ad6ec11c..47f4d160971e 100644
@@ -6268,7 +6268,7 @@ static void build_zonelists(pg_data_t *pgdat)
 		 */
 		if (node_distance(local_node, node) !=
 		    node_distance(local_node, prev_node))
-			node_load[node] = load;
+			node_load[node] += load;
 
 		node_order[nr_nodes++] = node;
 		prev_node = node;
Signature · 1 line
2.25.1
Acked-by on v1 2/2
#7 lore raw

Re: [FIX PATCH 2/2] mm/page_alloc: Use accumulated load when building node fallback list

Quoted 91 lines · Bharata B Rao

From: Krupa Ramakrishnan <[email protected]>

In build_zonelists(), when the fallback list is built for the nodes, the node load gets reinitialized during each iteration. This results in nodes with same distances occupying the same slot in different node fallback lists rather than appearing in the intended round- robin manner. This results in one node getting picked for allocation more compared to other nodes with the same distance.

As an example, consider a 4 node system with the following distance matrix.

Node 0 1 2 3
----------------

0    10 12 32 32
1    12 10 32 32
2    32 32 10 12
3    32 32 12 10

For this case, the node fallback list gets built like this:

Node Fallback list
---------------------
0 0 1 2 3
1 1 0 3 2
2 2 3 0 1
3 3 2 0 1 <-- Unexpected fallback order

In the fallback list for nodes 2 and 3, the nodes 0 and 1 appear in the same order which results in more allocations getting satisfied from node 0 compared to node 1.

The effect of this on remote memory bandwidth as seen by stream benchmark is shown below:

Case 1: Bandwidth from cores on nodes 2 & 3 to memory on nodes 0 & 1

	(numactl -m 0,1 ./stream_lowOverhead ... --cores <from 2, 3>)
Case 2: Bandwidth from cores on nodes 0 & 1 to memory on nodes 2 & 3
	(numactl -m 2,3 ./stream_lowOverhead ... --cores <from 0, 1>)

----------------------------------------

		BANDWIDTH (MB/s)
    TEST	Case 1		Case 2
----------------------------------------
    COPY	57479.6		110791.8
   SCALE	55372.9		105685.9
     ADD	50460.6		96734.2

TRIADD 50397.6 97119.1
----------------------------------------

The bandwidth drop in Case 1 occurs because most of the allocations get satisfied by node 0 as it appears first in the fallback order for both nodes 2 and 3.

This can be fixed by accumulating the node load in build_zonelists() rather than reinitializing it during each iteration. With this the nodes with the same distance rightly get assigned in the round robin manner. In fact this was how it was originally until the commit f0c0b2b808f2 ("change zonelist order: zonelist order selection logic") dropped the load accumulation and resorted to initializing the load during each iteration. While zonelist ordering was removed by commit c9bff3eebc09 ("mm, page_alloc: rip out ZONELIST_ORDER_ZONE"), the change to the node load accumulation in build_zonelists() remained. So essentially this patch reverts back to the accumulated node load logic.

After this fix, the fallback order gets built like this:

Node Fallback list
------------------

0    0 1 2 3
1    1 0 3 2
2    2 3 0 1
3    3 2 1 0 <-- Note the change here

The bandwidth in Case 1 improves and matches Case 2 as shown below.

----------------------------------------

		BANDWIDTH (MB/s)
    TEST	Case 1		Case 2
----------------------------------------
    COPY	110438.9	110107.2
   SCALE	105930.5	105817.5
     ADD	97005.1		96159.8

TRIADD 97441.5 96757.1
----------------------------------------

The correctness of the fallback list generation has been verified for the above node configuration where the node 3 starts as memory-less node and comes up online only during memory hotplug.

Quoted 8 lines

[[email protected]: Added changelog, review, test validation]

Fixes f0c0b2b808f2 ("change zonelist order: zonelist order selection

logic")

Minor not, The Fixes should be all one line even if it goes over the line length limit.

Otherwise, I can confirm that 2-socket Intel machines with SNC enabled suffer a similar problem -- the fallback lists for equal-distance nodes ends up overloading one node. Aside from the problems you mention, the overloaded node could trigger premature reclaim and inconsistent behaviour depending on where the task executes so

Signature · 2 lines
Mel Gorman
SUSE Labs
#8 lore raw

Re: [FIX PATCH 2/2] mm/page_alloc: Use accumulated load when building node fallback list

In reply to Bharata B Rao
Quoted 38 lines · Bharata B Rao

As an example, consider a 4 node system with the following distance matrix.

Node 0 1 2 3
----------------

0    10 12 32 32
1    12 10 32 32
2    32 32 10 12
3    32 32 12 10

For this case, the node fallback list gets built like this:

Node Fallback list
---------------------
0 0 1 2 3
1 1 0 3 2
2 2 3 0 1
3 3 2 0 1 <-- Unexpected fallback order

In the fallback list for nodes 2 and 3, the nodes 0 and 1 appear in the same order which results in more allocations getting satisfied from node 0 compared to node 1.

The effect of this on remote memory bandwidth as seen by stream benchmark is shown below:

Case 1: Bandwidth from cores on nodes 2 & 3 to memory on nodes 0 & 1

	(numactl -m 0,1 ./stream_lowOverhead ... --cores <from 2, 3>)
Case 2: Bandwidth from cores on nodes 0 & 1 to memory on nodes 2 & 3
	(numactl -m 2,3 ./stream_lowOverhead ... --cores <from 0, 1>)

----------------------------------------

		BANDWIDTH (MB/s)
    TEST	Case 1		Case 2
----------------------------------------
    COPY	57479.6		110791.8
   SCALE	55372.9		105685.9
     ADD	50460.6		96734.2
Quoted 6 lines

TRIADD 50397.6 97119.1
----------------------------------------

The bandwidth drop in Case 1 occurs because most of the allocations get satisfied by node 0 as it appears first in the fallback order for both nodes 2 and 3.

I am wondering what causes this performance drop here ? Would not the memory access latency be similar between {2, 3} ---> { 0 } and {2, 3} ---> { 1 }, given both these nodes {0, 1} have same distance from {2, 3} i.e 32 from the above distance matrix. Even if the preferred node order changes from { 0 } to { 1 } for the accessing node { 3 }, it should not change the latency as such.

Is the performance drop here, is caused by excessive allocation on node { 0 } resulting from page allocation latency instead.

#9 lore raw

RE: [FIX PATCH 2/2] mm/page_alloc: Use accumulated load when building node fallback list

[AMD Official Use Only]

The bandwidth is limited by underutilization of cross socket links and not the latency. Hotspotting on one node will not engage all hardware resources based on our routing protocol which results in the lower bandwidth. Distributing equally across nodes 0 and 1 will yield the best results as it stresses the full system capabilities.

Thanks
Krupa Ramakrishnan

-----Original Message-----
From: Anshuman Khandual <[email protected]>
Sent: 31 August, 2021 4:58
To: Rao, Bharata Bhasker <[email protected]>; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; Ramakrishnan, Krupa <[email protected]>; Srinivasan, Sadagopan <[email protected]>
Subject: Re: [FIX PATCH 2/2] mm/page_alloc: Use accumulated load when building node fallback list

[CAUTION: External Email]

Quoted 39 lines · Bharata B Rao

As an example, consider a 4 node system with the following distance matrix.

Node 0 1 2 3
----------------

0    10 12 32 32
1    12 10 32 32
2    32 32 10 12
3    32 32 12 10

For this case, the node fallback list gets built like this:

Node Fallback list
---------------------

0     0 1 2 3
1     1 0 3 2
2     2 3 0 1
3     3 2 0 1 <-- Unexpected fallback order

In the fallback list for nodes 2 and 3, the nodes 0 and 1 appear in the same order which results in more allocations getting satisfied from node 0 compared to node 1.

The effect of this on remote memory bandwidth as seen by stream benchmark is shown below:

Case 1: Bandwidth from cores on nodes 2 & 3 to memory on nodes 0 & 1

      (numactl -m 0,1 ./stream_lowOverhead ... --cores <from 2, 3>) 
Case 2: Bandwidth from cores on nodes 0 & 1 to memory on nodes 2 & 3
      (numactl -m 2,3 ./stream_lowOverhead ... --cores <from 0, 1>)

----------------------------------------
BANDWIDTH (MB/s)
TEST Case 1 Case 2
----------------------------------------

    COPY      57479.6         110791.8
   SCALE      55372.9         105685.9
     ADD      50460.6         96734.2
  TRIADD      50397.6         97119.1
Quoted 5 lines

----------------------------------------

The bandwidth drop in Case 1 occurs because most of the allocations get satisfied by node 0 as it appears first in the fallback order for both nodes 2 and 3.

I am wondering what causes this performance drop here ? Would not the memory access latency be similar between {2, 3} ---> { 0 } and {2, 3} ---> { 1 }, given both these nodes {0, 1} have same distance from {2, 3} i.e 32 from the above distance matrix. Even if the preferred node order changes from { 0 } to { 1 } for the accessing node { 3 }, it should not change the latency as such.

Is the performance drop here, is caused by excessive allocation on node { 0 } resulting from page allocation latency instead.

#10 lore raw

Re: [FIX PATCH 2/2] mm/page_alloc: Use accumulated load when building node fallback list

Quoted 3 lines · Krupa

[AMD Official Use Only]

The bandwidth is limited by underutilization of cross socket links and not the latency. Hotspotting on one node will not engage all hardware resources based on our routing protocol which results in the lower bandwidth. Distributing equally across nodes 0 and 1 will yield the best results as it stresses the full system capabilities.

Makes sense. Nonetheless this patch clearly solves a problem.

Quoted 63 lines

Thanks
Krupa Ramakrishnan

-----Original Message-----
From: Anshuman Khandual <[email protected]>
Sent: 31 August, 2021 4:58
To: Rao, Bharata Bhasker <[email protected]>; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; Ramakrishnan, Krupa <[email protected]>; Srinivasan, Sadagopan <[email protected]>
Subject: Re: [FIX PATCH 2/2] mm/page_alloc: Use accumulated load when building node fallback list

[CAUTION: External Email]

Quoted 39 lines · Bharata B Rao

As an example, consider a 4 node system with the following distance matrix.

Node 0 1 2 3
----------------

0    10 12 32 32
1    12 10 32 32
2    32 32 10 12
3    32 32 12 10

For this case, the node fallback list gets built like this:

Node Fallback list
---------------------

0     0 1 2 3
1     1 0 3 2
2     2 3 0 1
3     3 2 0 1 <-- Unexpected fallback order

In the fallback list for nodes 2 and 3, the nodes 0 and 1 appear in the same order which results in more allocations getting satisfied from node 0 compared to node 1.

The effect of this on remote memory bandwidth as seen by stream benchmark is shown below:

Case 1: Bandwidth from cores on nodes 2 & 3 to memory on nodes 0 & 1

      (numactl -m 0,1 ./stream_lowOverhead ... --cores <from 2, 3>) 
Case 2: Bandwidth from cores on nodes 0 & 1 to memory on nodes 2 & 3
      (numactl -m 2,3 ./stream_lowOverhead ... --cores <from 0, 1>)

----------------------------------------
BANDWIDTH (MB/s)
TEST Case 1 Case 2
----------------------------------------

    COPY      57479.6         110791.8
   SCALE      55372.9         105685.9
     ADD      50460.6         96734.2
  TRIADD      50397.6         97119.1
Quoted 5 lines

----------------------------------------

The bandwidth drop in Case 1 occurs because most of the allocations get satisfied by node 0 as it appears first in the fallback order for both nodes 2 and 3.

I am wondering what causes this performance drop here ? Would not the memory access latency be similar between {2, 3} ---> { 0 } and {2, 3} ---> { 1 }, given both these nodes {0, 1} have same distance from {2, 3} i.e 32 from the above distance matrix. Even if the preferred node order changes from { 0 } to { 1 } for the accessing node { 3 }, it should not change the latency as such.

Is the performance drop here, is caused by excessive allocation on node { 0 } resulting from page allocation latency instead.

Reviewed-by on v1 2/2
#11 lore raw

Re: [FIX PATCH 2/2] mm/page_alloc: Use accumulated load when building node fallback list

In reply to Bharata B Rao
Quoted 109 lines · Bharata B Rao

From: Krupa Ramakrishnan <[email protected]>

In build_zonelists(), when the fallback list is built for the nodes, the node load gets reinitialized during each iteration. This results in nodes with same distances occupying the same slot in different node fallback lists rather than appearing in the intended round- robin manner. This results in one node getting picked for allocation more compared to other nodes with the same distance.

As an example, consider a 4 node system with the following distance matrix.

Node 0 1 2 3
----------------

0    10 12 32 32
1    12 10 32 32
2    32 32 10 12
3    32 32 12 10

For this case, the node fallback list gets built like this:

Node Fallback list
---------------------
0 0 1 2 3
1 1 0 3 2
2 2 3 0 1
3 3 2 0 1 <-- Unexpected fallback order

In the fallback list for nodes 2 and 3, the nodes 0 and 1 appear in the same order which results in more allocations getting satisfied from node 0 compared to node 1.

The effect of this on remote memory bandwidth as seen by stream benchmark is shown below:

Case 1: Bandwidth from cores on nodes 2 & 3 to memory on nodes 0 & 1

	(numactl -m 0,1 ./stream_lowOverhead ... --cores <from 2, 3>)
Case 2: Bandwidth from cores on nodes 0 & 1 to memory on nodes 2 & 3
	(numactl -m 2,3 ./stream_lowOverhead ... --cores <from 0, 1>)

----------------------------------------

		BANDWIDTH (MB/s)
    TEST	Case 1		Case 2
----------------------------------------
    COPY	57479.6		110791.8
   SCALE	55372.9		105685.9
     ADD	50460.6		96734.2

TRIADD 50397.6 97119.1
----------------------------------------

The bandwidth drop in Case 1 occurs because most of the allocations get satisfied by node 0 as it appears first in the fallback order for both nodes 2 and 3.

This can be fixed by accumulating the node load in build_zonelists() rather than reinitializing it during each iteration. With this the nodes with the same distance rightly get assigned in the round robin manner. In fact this was how it was originally until the commit f0c0b2b808f2 ("change zonelist order: zonelist order selection logic") dropped the load accumulation and resorted to initializing the load during each iteration. While zonelist ordering was removed by commit c9bff3eebc09 ("mm, page_alloc: rip out ZONELIST_ORDER_ZONE"), the change to the node load accumulation in build_zonelists() remained. So essentially this patch reverts back to the accumulated node load logic.

After this fix, the fallback order gets built like this:

Node Fallback list
------------------

0    0 1 2 3
1    1 0 3 2
2    2 3 0 1
3    3 2 1 0 <-- Note the change here

The bandwidth in Case 1 improves and matches Case 2 as shown below.

----------------------------------------

		BANDWIDTH (MB/s)
    TEST	Case 1		Case 2
----------------------------------------
    COPY	110438.9	110107.2
   SCALE	105930.5	105817.5
     ADD	97005.1		96159.8

TRIADD 97441.5 96757.1
----------------------------------------

The correctness of the fallback list generation has been verified for the above node configuration where the node 3 starts as memory-less node and comes up online only during memory hotplug.

[[email protected]: Added changelog, review, test validation]

Fixes f0c0b2b808f2 ("change zonelist order: zonelist order selection

logic")

---
 mm/page_alloc.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/page_alloc.c +0 -0
index 22f7ad6ec11c..47f4d160971e 100644

mm/page_alloc.c static void build_zonelists(pg_data_t *pgdat) · line 6274

Quoted 8 lines
mm/page_alloc.c +1 -1
@@ -6268,7 +6268,7 @@ static void build_zonelists(pg_data_t *pgdat)
 		 */
 		if (node_distance(local_node, node) !=
 		    node_distance(local_node, prev_node))
-			node_load[node] = load;
+			node_load[node] += load;
 
 		node_order[nr_nodes++] = node;
 		prev_node = node;
#12 lore raw

Re: [FIX PATCH 2/2] mm/page_alloc: Use accumulated load when building node fallback list

In reply to Bharata B Rao
Quoted 19 lines · Bharata B Rao

From: Krupa Ramakrishnan <[email protected]>

In build_zonelists(), when the fallback list is built for the nodes, the node load gets reinitialized during each iteration. This results in nodes with same distances occupying the same slot in different node fallback lists rather than appearing in the intended round- robin manner. This results in one node getting picked for allocation more compared to other nodes with the same distance.

As an example, consider a 4 node system with the following distance matrix.

Node 0 1 2 3
----------------

0    10 12 32 32
1    12 10 32 32
2    32 32 10 12
3    32 32 12 10
Quoted 8 lines

For this case, the node fallback list gets built like this:

Node Fallback list
---------------------
0 0 1 2 3
1 1 0 3 2
2 2 3 0 1
3 3 2 0 1 <-- Unexpected fallback order

FWIW, for a dual-socket 8 node system with the following distance matrix,

node 0 1 2 3 4 5 6 7

  0:  10  12  12  12  32  32  32  32
  1:  12  10  12  12  32  32  32  32
  2:  12  12  10  12  32  32  32  32
  3:  12  12  12  10  32  32  32  32
  4:  32  32  32  32  10  12  12  12
  5:  32  32  32  32  12  10  12  12
  6:  32  32  32  32  12  12  10  12
  7:  32  32  32  32  12  12  12  10

the fallback list looks like this:

Before
=======
Fallback order for Node 0: 0 1 2 3 4 5 6 7 Fallback order for Node 1: 1 2 3 0 5 6 7 4 Fallback order for Node 2: 2 3 0 1 6 7 4 5 Fallback order for Node 3: 3 0 1 2 7 4 5 6 Fallback order for Node 4: 4 5 6 7 0 1 2 3 Fallback order for Node 5: 5 6 7 4 0 1 2 3 Fallback order for Node 6: 6 7 4 5 0 1 2 3 Fallback order for Node 7: 7 4 5 6 0 1 2 3

After the fix
==============
Fallback order for Node 0: 0 1 2 3 4 5 6 7 Fallback order for Node 1: 1 2 3 0 5 6 7 4 Fallback order for Node 2: 2 3 0 1 6 7 4 5 Fallback order for Node 3: 3 0 1 2 7 4 5 6 Fallback order for Node 4: 4 5 6 7 0 1 2 3 Fallback order for Node 5: 5 6 7 4 1 2 3 0 Fallback order for Node 6: 6 7 4 5 2 3 0 1 Fallback order for Node 7: 7 4 5 6 3 0 1 2

So the problem becomes more pronounced for bigger NUMA systems.

Regards,
Bharata.