Skip to contentSkip to conversation

sched,numa: document and fix numa_preferred_nid setting

PATCH In review Rik van Riel 3 participants 14 messages Jun 16, 2015 – Jun 22, 2015 Scheduler

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched,numa: document and fix numa_preferred_nid setting

There are two places where the numa balancing code sets a task's numa_preferred_nid.

The primary location is task_numa_placement(), where the kernel examines the NUMA fault statistics to determine the location where most of the memory that the task (or numa_group) accesses is.

The second location is only used for large workloads, where a numa_group has enough tasks that the tasks are spread out over several NUMA nodes, and multiple nodes are in the numa group's active_nodes mask.

In order to allow those workloads to settle down, we pretend that any node inside the numa_group's active_nodes mask is the task's new preferred node. This dissuades task_numa_fault() from continuously retrying to migrate the task to the group's preferred node, and allows a multi-node workload to settle down. This in turn improves locality of private faults inside a numa group.

---
 kernel/sched/fair.c | 23 +++++++++++++++--------
 1 file changed, 15 insertions(+), 8 deletions(-)
kernel/sched/fair.c +15 -8
index c2980e8733bc..54bb57f09e75 100644
@@ -1485,7 +1485,12 @@ static int task_numa_migrate(struct task_struct *p)
 				groupweight = group_weight(p, env.src_nid, dist);
 			}
 
-			/* Only consider nodes where both task and groups benefit */
+			/*
+			 * Only consider nodes where placement is better for
+			 * either the group (help large workloads converge),
+			 * or the task (placement of tasks within a numa group,
+			 * and single threaded processes).
+			 */
 			taskimp = task_weight(p, nid, dist) - taskweight;
 			groupimp = group_weight(p, nid, dist) - groupweight;
 			if (taskimp < 0 && groupimp < 0)
@@ -1499,12 +1504,14 @@ static int task_numa_migrate(struct task_struct *p)
 	}
 
 	/*
-	 * If the task is part of a workload that spans multiple NUMA nodes,
-	 * and is migrating into one of the workload's active nodes, remember
-	 * this node as the task's preferred numa node, so the workload can
-	 * settle down.
-	 * A task that migrated to a second choice node will be better off
-	 * trying for a better one later. Do not set the preferred node here.
+	 * The primary place for setting a task's numa_preferred_nid is in
+	 * task_numa_placement(). If a task is moved to a sub-optimal node,
+	 * leave numa_preferred_nid alone, so task_numa_fault() will retry
+	 * migrating the task to where it really belongs.
+	 * The exception is a task that belongs to a large numa_group, which
+	 * spans multiple NUMA nodes. If that task migrates into one of the
+	 * workload's active nodes, remember that node as the task's
+	 * numa_preferred_nid, so the workload can settle down.
 	 */
 	if (p->numa_group) {
 		if (env.best_cpu == -1)
@@ -1513,7 +1520,7 @@ static int task_numa_migrate(struct task_struct *p)
 			nid = env.dst_nid;
 
 		if (node_isset(nid, p->numa_group->active_nodes))
-			sched_setnuma(p, env.dst_nid);
+			sched_setnuma(p, nid);
 	}
 
 	/* No better CPU than the current one was found. */
#1 lore raw

kernel/sched/fair.c static int task_numa_migrate(struct task_struct *p) · line 1526

Quoted 12 lines
kernel/sched/fair.c +1 -1
 	if (p->numa_group) {
 		if (env.best_cpu == -1)
@@ -1513,7 +1520,7 @@ static int task_numa_migrate(struct task_struct *p)
 			nid = env.dst_nid;
 
 		if (node_isset(nid, p->numa_group->active_nodes))
-			sched_setnuma(p, env.dst_nid);
+			sched_setnuma(p, nid);
 	}
 
 	/* No better CPU than the current one was found. */

Overall this patch does seem to produce better results. However numa02 gets affected -vely.

KernelVersion: 4.1.0-rc7-tip
Testcase: Min Max Avg StdDev
elapsed_numa01: 858.85 949.18 915.64 33.06
elapsed_numa02: 23.09 29.89 26.43 2.18
Testcase: Min Max Avg StdDev
system_numa01: 1516.72 1855.08 1686.24 113.95
system_numa02: 63.69 79.06 70.35 5.87

	Testcase:         Min         Max         Avg      StdDev
     user_numa01:    73284.76    80818.21    78060.88     2773.60
     user_numa02:     1690.18     2071.07     1821.64      140.25
	Testcase:         Min         Max         Avg      StdDev
    total_numa01:    74801.50    82572.60    79747.12     2875.61
    total_numa02:     1753.87     2142.77     1891.99      143.59

KernelVersion: 4.1.0-rc7-tip + your patch

Testcase: Min Max Avg StdDev %Change
elapsed_numa01: 665.26 877.47 776.77 79.23 15.83%
elapsed_numa02: 24.59 31.30 28.17 2.48 -5.56%
Testcase: Min Max Avg StdDev %Change
system_numa01: 659.57 1220.99 942.36 234.92 60.92%
system_numa02: 44.62 86.01 64.64 14.24 6.64%

	Testcase:         Min         Max         Avg      StdDev     %Change
     user_numa01:    56280.95    75908.81    64993.57     7764.30      17.21%
     user_numa02:     1790.35     2155.02     1916.12      132.57      -4.38%
	Testcase:         Min         Max         Avg      StdDev     %Change
    total_numa01:    56940.50    77128.20    65935.92     7993.49      17.91%
    total_numa02:     1834.97     2227.03     1980.76      136.51      -3.99%
Signature · 2 lines
Thanks and Regards
Srikar Dronamraju
#2 lore raw
Quoted 10 lines · Srikar Dronamraju
Quoted 10 lines
kernel/sched/fair.c +1 -1
 	if (p->numa_group) {
 		if (env.best_cpu == -1)
@@ -1513,7 +1520,7 @@ static int task_numa_migrate(struct task_struct *p)
 			nid = env.dst_nid;
 
 		if (node_isset(nid, p->numa_group->active_nodes))
-			sched_setnuma(p, env.dst_nid);
+			sched_setnuma(p, nid);
 	}
 
Quoted 5 lines

kernel/sched/fair.c static int task_numa_migrate(struct task_struct *p) · line 1526

Quoted 2 lines
kernel/sched/fair.c +0 -0
 	/* No better CPU than the current one was found. */

Overall this patch does seem to produce better results. However numa02 gets affected -vely.

OK, that is kind of expected.

The way numa02 runs means that we are essentially guaranteed that, on a two node system, both nodes end up in the numa_group's active_mask.

What the above change does is slow down migration if a task ends up in a NUMA node in p->numa_group->active_nodes.

This is necessary if a very large workload has already converged on a set of NUMA nodes, but it does slow down convergence for such workloads.

I can't think of any obvious way to both slow down movement once things have converged, yet keep speedy movement of tasks when they have not yet converged.

It is worth noting that all the numa01 and numa02 benchmarks measure is the speed at which the workloads converge. It does not measure the overhead of making things converge, or how fast an actual workload runs (NUMA locality benefit, minus NUMA placement overhead).

Quoted 29 lines

KernelVersion: 4.1.0-rc7-tip
Testcase: Min Max Avg StdDev
elapsed_numa01: 858.85 949.18 915.64 33.06
elapsed_numa02: 23.09 29.89 26.43 2.18
Testcase: Min Max Avg StdDev
system_numa01: 1516.72 1855.08 1686.24 113.95
system_numa02: 63.69 79.06 70.35 5.87

	Testcase:         Min         Max         Avg      StdDev
     user_numa01:    73284.76    80818.21    78060.88     2773.60
     user_numa02:     1690.18     2071.07     1821.64      140.25
	Testcase:         Min         Max         Avg      StdDev
    total_numa01:    74801.50    82572.60    79747.12     2875.61
    total_numa02:     1753.87     2142.77     1891.99      143.59

KernelVersion: 4.1.0-rc7-tip + your patch

Testcase: Min Max Avg StdDev %Change
elapsed_numa01: 665.26 877.47 776.77 79.23 15.83%
elapsed_numa02: 24.59 31.30 28.17 2.48 -5.56%
Testcase: Min Max Avg StdDev %Change
system_numa01: 659.57 1220.99 942.36 234.92 60.92%
system_numa02: 44.62 86.01 64.64 14.24 6.64%

	Testcase:         Min         Max         Avg      StdDev     %Change
     user_numa01:    56280.95    75908.81    64993.57     7764.30      17.21%
     user_numa02:     1790.35     2155.02     1916.12      132.57      -4.38%
	Testcase:         Min         Max         Avg      StdDev     %Change
    total_numa01:    56940.50    77128.20    65935.92     7993.49      17.91%
    total_numa02:     1834.97     2227.03     1980.76      136.51      -3.99%
#3 lore raw
Quoted 11 lines · Rik van Riel
Quoted 4 lines
Quoted 1 line

Overall this patch does seem to produce better results. However numa02 gets affected -vely.

OK, that is kind of expected.

The way numa02 runs means that we are essentially guaranteed that, on a two node system, both nodes end up in the numa_group's active_mask.

Just to add this was on a 4 node machine.

Signature · 2 lines
Thanks and Regards
Srikar Dronamraju
#4 lore raw
Quoted 7 lines · Srikar Dronamraju
Quoted 5 lines · Rik van Riel
Quoted 4 lines
Quoted 1 line

Overall this patch does seem to produce better results. However numa02 gets affected -vely.

Quoted 8 lines
Quoted 6 lines

OK, that is kind of expected.

The way numa02 runs means that we are essentially guaranteed that, on a two node system, both nodes end up in the numa_group's active_mask.

Just to add this was on a 4 node machine.

OK, so we are looking at two multi-threaded processes on a 4 node system, and waiting for them to converge?

It may make sense to add my patch in with your patch 1/4 from last week, as well as the correct part of your patch 4/4, and see how they all work together.

#5 lore raw
Quoted 8 lines

OK, so we are looking at two multi-threaded processes on a 4 node system, and waiting for them to converge?

It may make sense to add my patch in with your patch 1/4 from last week, as well as the correct part of your patch 4/4, and see how they all work together. --

Okay, I will do the needful and come back to you.

Signature · 2 lines
Thanks and Regards
Srikar Dronamraju
#6 lore raw
In reply to Rik van Riel
Quoted 8 lines

OK, so we are looking at two multi-threaded processes on a 4 node system, and waiting for them to converge?

It may make sense to add my patch in with your patch 1/4 from last week, as well as the correct part of your patch 4/4, and see how they all work together.

Tested specjbb and autonumabenchmark on 4 kernels.

Plain 4.1.0-rc7-tip (i)
tip + only Rik's patch (ii)
tip + Rik's ++ (iii)
tip + Srikar's ++ (iv)

(i) = Plain 4.1.0-rc7-tip = tip = 4.1.0-rc7 (b7ca96b)

(ii) = tip + only Rik's patch = (i) + Rik's fix numa_preferred_nid setting

(iii) = tip + Rik's ++ (iii) = (ii) + Srikar's numa hotness + correct nid for evaluating task weight

(iv) = tip + Srikar's ++ (iv) = (i) + Srikar's numa hotness + correct nid for evaluating task weight +
numa_has_capacity fix + always update preferred node

Plain 4.1.0-rc7-tip (i)

		Testcase:         Min         Max         Avg      StdDev
	  elapsed_numa01:      858.85      949.18      915.64       33.06
	  elapsed_numa02:       23.09       29.89       26.43        2.18
		Testcase:         Min         Max         Avg      StdDev
	   system_numa01:     1516.72     1855.08     1686.24      113.95
	   system_numa02:       63.69       79.06       70.35        5.87
		Testcase:         Min         Max         Avg      StdDev
	     user_numa01:    73284.76    80818.21    78060.88     2773.60
	     user_numa02:     1690.18     2071.07     1821.64      140.25
		Testcase:         Min         Max         Avg      StdDev
	    total_numa01:    74801.50    82572.60    79747.12     2875.61
	    total_numa02:     1753.87     2142.77     1891.99      143.59

tip + only Rik's patch (ii)

		Testcase:         Min         Max         Avg      StdDev     %Change
	  elapsed_numa01:      665.26      877.47      776.77       79.23      15.83%
	  elapsed_numa02:       24.59       31.30       28.17        2.48      -5.56%
		Testcase:         Min         Max         Avg      StdDev     %Change
	   system_numa01:      659.57     1220.99      942.36      234.92      60.92%
	   system_numa02:       44.62       86.01       64.64       14.24       6.64%
		Testcase:         Min         Max         Avg      StdDev     %Change
	     user_numa01:    56280.95    75908.81    64993.57     7764.30      17.21%
	     user_numa02:     1790.35     2155.02     1916.12      132.57      -4.38%
		Testcase:         Min         Max         Avg      StdDev     %Change
	    total_numa01:    56940.50    77128.20    65935.92     7993.49      17.91%
	    total_numa02:     1834.97     2227.03     1980.76      136.51      -3.99%

tip + Rik's ++ (iii)

		Testcase:         Min         Max         Avg      StdDev     %Change
	  elapsed_numa01:      630.60      860.06      760.07       74.33      18.09%
	  elapsed_numa02:       21.92       34.42       27.72        4.49      -3.75%
		Testcase:         Min         Max         Avg      StdDev     %Change
	   system_numa01:      474.31     1379.49      870.12      296.35      59.16%
	   system_numa02:       63.74      120.25       86.69       20.69     -13.59%
		Testcase:         Min         Max         Avg      StdDev     %Change
	     user_numa01:    53004.03    68125.84    61697.01     5011.38      24.02%
	     user_numa02:     1650.82     2278.71     1941.26      224.59      -5.25%
		Testcase:         Min         Max         Avg      StdDev     %Change
	    total_numa01:    53478.30    69505.30    62567.12     5288.18      24.72%
	    total_numa02:     1714.56     2398.96     2027.95      238.08      -5.67%

tip + Srikar's ++ (iv)

		Testcase:         Min         Max         Avg      StdDev     %Change
	  elapsed_numa01:      690.74      919.49      782.67       78.51      14.46%
	  elapsed_numa02:       21.78       29.57       26.02        2.65       1.39%
		Testcase:         Min         Max         Avg      StdDev     %Change
	   system_numa01:      659.12     1041.19      870.15      143.13      78.38%
	   system_numa02:       52.20       78.73       64.18       11.28       7.84%
		Testcase:         Min         Max         Avg      StdDev     %Change
	     user_numa01:    56410.39    71492.31    62514.78     5444.90      21.75%
	     user_numa02:     1594.27     1934.40     1754.37      126.41       3.48%
		Testcase:         Min         Max         Avg      StdDev     %Change
	    total_numa01:    57069.50    72509.90    63384.94     5567.71      22.57%
	    total_numa02:     1647.85     2010.87     1818.55      136.88       3.65%

5 interations of Specjbb on 4 node, 24 core powerpc machine. Ran 1 instance per system.

For specjbb (higher bops per JVM is better)

Plain 4.1.0-rc7-tip (i)

	  Metric:         Min         Max         Avg      StdDev
      bopsperJVM:   265519.00   272466.00   269377.80     2391.04

tip + only Rik's patch (ii)

	  Metric:         Min         Max         Avg      StdDev     %Change
      bopsperJVM:   263393.00   269660.00   266920.20     2792.07      -0.91%

tip + Rik's ++ (iii)

	  Metric:         Min         Max         Avg      StdDev     %Change
      bopsperJVM:   264298.00   271236.00   266818.20     2579.62      -0.94%

tip + Srikar's ++ (iv)

	  Metric:         Min         Max         Avg      StdDev     %Change
      bopsperJVM:   266774.00   272434.00   269839.60 	  2083.19      +0.17%

So fix for numa_has_capacity and always setting preferred node based on fault stats seems to help autonuma and specjbb.

Signature · 6 lines
Thanks and Regards
Srikar Dronamraju

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/
#7 lore raw
Quoted 11 lines · Srikar Dronamraju
Quoted 8 lines

OK, so we are looking at two multi-threaded processes on a 4 node system, and waiting for them to converge?

It may make sense to add my patch in with your patch 1/4 from last week, as well as the correct part of your patch 4/4, and see how they all work together.

Tested specjbb and autonumabenchmark on 4 kernels.

Quoted 7 lines

Plain 4.1.0-rc7-tip (i)
tip + only Rik's patch (ii)
tip + Rik's ++ (iii)
tip + Srikar's ++ (iv)

5 interations of Specjbb on 4 node, 24 core powerpc machine. Ran 1 instance per system.

Would you happen to have 2 instance and 4 instance SPECjbb numbers, too? The single instance numbers seem to be within the margin of error, but I would expect multi-instance numbers to show more dramatic changes, due to changes in how workloads converge...

Those behave very differently from single instance, especially with the "always set the preferred_nid, even if we moved the task to a node we do NOT prefer" patch...

It would be good to understand the behaviour of these patches under more circumstances.

Signature · 4 lines
All rights reversed
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/
#8 lore raw
Quoted 7 lines

Would you happen to have 2 instance and 4 instance SPECjbb numbers, too? The single instance numbers seem to be within the margin of error, but I would expect multi-instance numbers to show more dramatic changes, due to changes in how workloads converge...

Quoted 6 lines

Those behave very differently from single instance, especially with the "always set the preferred_nid, even if we moved the task to a node we do NOT prefer" patch...

It would be good to understand the behaviour of these patches under more circumstances.

Here are specjbb2005 numbers with 1 JVM per System, 2 JVMs per System and 4 JVMs per System.

Plain 4.1.0-rc7-tip (i)
tip + Rik's ++ (ii)
tip + Srikar's ++ (iii)
tip + Srikar's + Modified Rik's patch (iv)

(i) = Plain 4.1.0-rc7-tip = tip = 4.1.0-rc7 (b7ca96b)

(ii) = tip + only Rik's suggested patches = (i) + Rik's fix numa_preferred_nid setting
+ Srikar's numa hotness + correct nid for evaluating task weight

(iii) = tip + Srikar's ++ (iii) = (i) + Srikar's numa hotness + correct nid for evaluating
task weight + numa_has_capacity fix + always update preferred node

(iv) = tip + Srikar's ++ (iv) = (i) + Srikar's numa hotness + correct nid for evaluating
task weight + numa_has_capacity fix + Rik's modified patch.
(Rik's modified patch == I removed node_isset check before setting nid as the preferred node)

jbb2005_1JVMperSYSTEM
Plain 4.1.0-rc7-tip (i)

		  Metric:         Min         Max         Avg      StdDev     %Change
	      bopsperJVM:   265519.00   272466.00   269377.80     2391.04

tip + Rik's ++ (ii)

	      bopsperJVM:   264298.00   271236.00   266818.20     2579.62      -0.94%

tip + Srikar's ++ (iii)

	      bopsperJVM:   266774.00   272434.00   269839.60     2083.19       0.17%

tip + Srikar's + Rik's (iv)

	      bopsperJVM:   265037.00   274419.00   269280.00     3146.74      -0.04%

jbb2005_2JVMperSYSTEM
Plain 4.1.0-rc7-tip (i)

		  Metric:         Min         Max         Avg      StdDev     %Change
	      bopsperJVM:   269575.00   288495.00   279910.80     6151.49

tip + Srikar's ++ (iii)

	      bopsperJVM:   278810.00   287706.00   282514.00     2946.37      0.90%

tip + Rik's ++ (ii)

	      bopsperJVM:   286785.00   289515.00   288311.80     1206.66      2.90%

tip + Srikar's + Rik's (iv)

	      bopsperJVM:   283295.00   293466.00   287848.80     3427.06      2.70%

jbb2005_4JVMperSYSTEM
Plain 4.1.0-rc7-tip (i)

		  Metric:         Min         Max         Avg      StdDev     %Change
	      bopsperJVM:   248392.00   263826.00   257263.20     5946.44

tip + Rik's ++ (ii)

	      bopsperJVM:   257057.00   260303.00   258819.00     1234.46      0.60%

tip + Srikar's ++ (iii)

	      bopsperJVM:   252968.00   262006.00   257321.80     3131.00      0.02%

tip + Srikar's + Rik's (iv)

	      bopsperJVM:   257063.00   266196.00   262547.80     3099.57      1.99%

Summary:
While Rik's suggested patchset performs the best in 2 JVM case and numa01. A modified version of his patch, provides good performance in 2 JVM, 4 JVM cases and numa01. However these two patchsets dont regress in numa02 (probably a little less with modified patch)

Signature · 6 lines
Thanks and Regards
Srikar Dronamraju

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/
#9 lore raw
In reply to Srikar Dronamraju

Updated autonumabenchmark numbers

Plain 4.1.0-rc7-tip (i)

		Testcase:         Min         Max         Avg      StdDev
	  elapsed_numa01:      858.85      949.18      915.64       33.06
	  elapsed_numa02:       23.09       29.89       26.43        2.18
		Testcase:         Min         Max         Avg      StdDev
	   system_numa01:     1516.72     1855.08     1686.24      113.95
	   system_numa02:       63.69       79.06       70.35        5.87
		Testcase:         Min         Max         Avg      StdDev
	     user_numa01:    73284.76    80818.21    78060.88     2773.60
	     user_numa02:     1690.18     2071.07     1821.64      140.25
		Testcase:         Min         Max         Avg      StdDev
	    total_numa01:    74801.50    82572.60    79747.12     2875.61
	    total_numa02:     1753.87     2142.77     1891.99      143.59

tip + Rik's ++ (ii)

		Testcase:         Min         Max         Avg      StdDev     %Change
	  elapsed_numa01:      630.60      860.06      760.07       74.33      18.09%
	  elapsed_numa02:       21.92       34.42       27.72        4.49      -3.75%
		Testcase:         Min         Max         Avg      StdDev     %Change
	   system_numa01:      474.31     1379.49      870.12      296.35      59.16%
	   system_numa02:       63.74      120.25       86.69       20.69     -13.59%
		Testcase:         Min         Max         Avg      StdDev     %Change
	     user_numa01:    53004.03    68125.84    61697.01     5011.38      24.02%
	     user_numa02:     1650.82     2278.71     1941.26      224.59      -5.25%
		Testcase:         Min         Max         Avg      StdDev     %Change
	    total_numa01:    53478.30    69505.30    62567.12     5288.18      24.72%
	    total_numa02:     1714.56     2398.96     2027.95      238.08      -5.67%

tip + Srikar's ++ (iii)

		Testcase:         Min         Max         Avg      StdDev     %Change
	  elapsed_numa01:      690.74      919.49      782.67       78.51      14.46%
	  elapsed_numa02:       21.78       29.57       26.02        2.65       1.39%
		Testcase:         Min         Max         Avg      StdDev     %Change
	   system_numa01:      659.12     1041.19      870.15      143.13      78.38%
	   system_numa02:       52.20       78.73       64.18       11.28       7.84%
		Testcase:         Min         Max         Avg      StdDev     %Change
	     user_numa01:    56410.39    71492.31    62514.78     5444.90      21.75%
	     user_numa02:     1594.27     1934.40     1754.37      126.41       3.48%
		Testcase:         Min         Max         Avg      StdDev     %Change
	    total_numa01:    57069.50    72509.90    63384.94     5567.71      22.57%
	    total_numa02:     1647.85     2010.87     1818.55      136.88       3.65%

tip + Srikar's + Modified Rik's patch (iv)

		Testcase:         Min         Max         Avg      StdDev     %Change
	  elapsed_numa01:      674.72      815.10      746.50       51.81      20.75%
	  elapsed_numa02:       21.02       34.57       27.58        4.63      -3.33%
		Testcase:         Min         Max         Avg      StdDev     %Change
	   system_numa01:      726.19     1099.16      879.36      141.02      73.41%
	   system_numa02:       33.99       72.99       58.89       14.26      15.70%
		Testcase:         Min         Max         Avg      StdDev     %Change
	     user_numa01:    56350.75    67318.49    61998.95     4696.34      23.86%
	     user_numa02:     1518.66     2301.80     1882.96      261.18      -2.66%
		Testcase:         Min         Max         Avg      StdDev     %Change
	    total_numa01:    57076.90    68417.70    62878.32     4826.93      24.66%
	    total_numa02:     1552.65     2374.79     1941.85      274.28      -2.10%

Plain 4.1.0-rc7-tip (i)
tip + Rik's ++ (ii)
tip + Srikar's ++ (iii)
tip + Srikar's + Modified Rik's patch (iv)

(i) = Plain 4.1.0-rc7-tip = tip = 4.1.0-rc7 (b7ca96b)

(ii) = tip + only Rik's suggested patches = (i) + Rik's fix numa_preferred_nid setting
+ Srikar's numa hotness + correct nid for evaluating task weight

(iii) = tip + Srikar's ++ (iii) = (i) + Srikar's numa hotness + correct nid for evaluating
task weight + numa_has_capacity fix + always update preferred node

(iv) = tip + Srikar's ++ (iv) = (i) + Srikar's numa hotness + correct nid for evaluating
task weight + numa_has_capacity fix + Rik's modified patch.
(Rik's modified patch == I removed node_isset check before setting nid as the preferred node)

Signature · 6 lines
Thanks and Regards
Srikar Dronamraju

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/
#10 lore raw
In reply to Srikar Dronamraju
Quoted 10 lines · Srikar Dronamraju
Quoted 10 lines
kernel/sched/fair.c +1 -1
 	if (p->numa_group) {
 		if (env.best_cpu == -1)
@@ -1513,7 +1520,7 @@ static int task_numa_migrate(struct task_struct *p)
 			nid = env.dst_nid;
 
 		if (node_isset(nid, p->numa_group->active_nodes))
-			sched_setnuma(p, env.dst_nid);
+			sched_setnuma(p, nid);
 	}
 
Quoted 5 lines

kernel/sched/fair.c static int task_numa_migrate(struct task_struct *p) · line 1526

Quoted 2 lines
kernel/sched/fair.c +0 -0
 	/* No better CPU than the current one was found. */

Overall this patch does seem to produce better results. However numa02 gets affected -vely.

Huh?

numa02 is the more important benchmark of the two. 'numa01' is a conflicting workload that is a lot more sensitive to balancing details - while 'numa02' is a nicely partitioned workload that should converge as fast as possible.

So if numa02 got worse then it's a bad change.

Thanks,

Ingo

#11 lore raw
Quoted 18 lines · Ingo Molnar
Quoted 10 lines · Srikar Dronamraju
Quoted 10 lines
kernel/sched/fair.c +1 -1
 	if (p->numa_group) {
 		if (env.best_cpu == -1)
@@ -1513,7 +1520,7 @@ static int task_numa_migrate(struct task_struct *p)
 			nid = env.dst_nid;
 
 		if (node_isset(nid, p->numa_group->active_nodes))
-			sched_setnuma(p, env.dst_nid);
+			sched_setnuma(p, nid);
 	}
 
Quoted 5 lines

kernel/sched/fair.c static int task_numa_migrate(struct task_struct *p) · line 1526

Quoted 2 lines
kernel/sched/fair.c +0 -0
 	/* No better CPU than the current one was found. */

Overall this patch does seem to produce better results. However numa02 gets affected -vely.

Quoted 7 lines

Huh?

numa02 is the more important benchmark of the two. 'numa01' is a conflicting workload that is a lot more sensitive to balancing details - while 'numa02' is a nicely partitioned workload that should converge as fast as possible.

So if numa02 got worse then it's a bad change.

It slows down convergence.

However, for a benchmark that spends much of its time actually doing something with the memory it accesses, after things have converged, slowing down task movement after convergence may be a good thing.

It will be good to get SPECjbb2005 numbers, both with 1 instance, 2 instances, and 4 instances on the same 4 node system.

If it turns out that single instance and two instance workloads do not benefit from trying to slow down task movement after the workloads have converged, we can change the code above, and simply retry task migration much, much more often after the workload has converged onto several nodes.

#12 lore raw
In reply to Rik van Riel
Quoted 10 lines
kernel/sched/fair.c +5 -0
+	 * migrating the task to where it really belongs.
+	 * The exception is a task that belongs to a large numa_group, which
+	 * spans multiple NUMA nodes. If that task migrates into one of the
+	 * workload's active nodes, remember that node as the task's
+	 * numa_preferred_nid, so the workload can settle down.
 	 */
 	if (p->numa_group) {
 		if (env.best_cpu == -1)

kernel/sched/fair.c static int task_numa_migrate(struct task_struct *p) · line 1526

Quoted 8 lines
kernel/sched/fair.c +1 -1
@@ -1513,7 +1520,7 @@ static int task_numa_migrate(struct task_struct *p)
 			nid = env.dst_nid;
 
 		if (node_isset(nid, p->numa_group->active_nodes))
-			sched_setnuma(p, env.dst_nid);
+			sched_setnuma(p, nid);
 	}
 
 	/* No better CPU than the current one was found. */

When I refer to the Modified Rik's patch, I mean to remove the node_isset() check before the sched_setnuma. With that change, we kind of reduce the numa02 and 1JVMper System regression while getting as good numbers as Rik's patch with 2JVM and 4JVM per System.

The idea behind removing the node_isset check is:
node_isset is mostly used to track mem movement to nodes where cpus are running and not vice versa. This is as per comment in update_numa_active_node_mask. There could be a sitation where task memory is all in a node and the node has capacity to accomodate but no tasks associated with the task have run enuf on that node. In such a case, we shouldnt be ruling out migrating the task to the node.

Signature · 6 lines
Thanks and Regards
Srikar Dronamraju

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/
#13 lore raw
Quoted 24 lines · Srikar Dronamraju
Quoted 10 lines
kernel/sched/fair.c +5 -0
+	 * migrating the task to where it really belongs.
+	 * The exception is a task that belongs to a large numa_group, which
+	 * spans multiple NUMA nodes. If that task migrates into one of the
+	 * workload's active nodes, remember that node as the task's
+	 * numa_preferred_nid, so the workload can settle down.
 	 */
 	if (p->numa_group) {
 		if (env.best_cpu == -1)

kernel/sched/fair.c static int task_numa_migrate(struct task_struct *p) · line 1526

Quoted 8 lines
kernel/sched/fair.c +1 -1
@@ -1513,7 +1520,7 @@ static int task_numa_migrate(struct task_struct *p)
 			nid = env.dst_nid;
 
 		if (node_isset(nid, p->numa_group->active_nodes))
-			sched_setnuma(p, env.dst_nid);
+			sched_setnuma(p, nid);
 	}
 
 	/* No better CPU than the current one was found. */

When I refer to the Modified Rik's patch, I mean to remove the node_isset() check before the sched_setnuma. With that change, we kind of reduce the numa02 and 1JVMper System regression while getting as good numbers as Rik's patch with 2JVM and 4JVM per System.

Quoted 7 lines

The idea behind removing the node_isset check is:
node_isset is mostly used to track mem movement to nodes where cpus are running and not vice versa. This is as per comment in update_numa_active_node_mask. There could be a sitation where task memory is all in a node and the node has capacity to accomodate but no tasks associated with the task have run enuf on that node. In such a case, we shouldnt be ruling out migrating the task to the node.

That is a good point.

However, if overriding the preferred_nid that task_numa_placement identified is a good idea in task_numa_migrate, would it also be a good idea for tasks that are NOT part of a numa group?

What are the consequences of never setting preferred_nid from
task_numa_migrate? (we would try to migrate the task to a
better node more frequently)

What are the consequences of always setting preferred_nid from
task_numa_migrate? (we would only try migrating the task once,
and it could get stuck in a sub-optimal location)

The patch seems to work, but I do not understand why, and would like to know your ideas on why you think the patch works.

I am really not looking forward to the idea of maintaining code that nobody understands...

Signature · 4 lines
All rights reversed
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/