numa,sched: resolve conflict between load balancing and NUMA balancing

A previous attempt to resolve a major conflict between load balancing and NUMA balancing, changeset 095bebf61a46 ("sched/numa: Do not move past the balance point if unbalanced"), introduced its own problems.

Revert that changeset, and introduce a new fix, which actually seems to resolve the issues observed in Jirka's tests.

A test where the main thread creates a large memory area, and spawns a worker thread to iterate over the memory (placed on another node by select_task_rq_fair), after which the main thread goes to sleep and waits for the worker thread to loop over all the memory now sees the worker thread migrated to where the memory is, instead of having all the memory migrated over like before.

Jirka has run a number of performance tests on several systems:
single instance SpecJBB 2005 performance is 7-15% higher on a 4 node system, with higher gains on systems with more cores per socket. Multi-instance SpecJBB 2005 (one per node), linpack, and stream see little or no changes with the revert of 095bebf61a46 and this patch.

From: Rik van Riel <[email protected]>

Commit 095bebf61a46 ("sched/numa: Do not move past the balance point if unbalanced") broke convergence of workloads with just one runnable thread, by making it impossible for the one runnable thread on the system to move from one NUMA node to another.

Instead, the thread would remain where it was, and pull all the memory across to its location, which is much slower than just migrating the thread to where the memory is.

The next patch has a better fix for the issue that 095bebf61a46 tried to address.

---
 kernel/sched/fair.c | 41 +++++++++++++++--------------------------
 1 file changed, 15 insertions(+), 26 deletions(-)

index ffeaa4105e48..c47bf0dffb34 100644
@@ -1198,11 +1198,9 @@ static void task_numa_assign(struct task_numa_env *env,
 static bool load_too_imbalanced(long src_load, long dst_load,
 				struct task_numa_env *env)
 {
+	long imb, old_imb;
+	long orig_src_load, orig_dst_load;
 	long src_capacity, dst_capacity;
-	long orig_src_load;
-	long load_a, load_b;
-	long moved_load;
-	long imb;
 
 	/*
 	 * The load is corrected for the CPU capacity available on each node.

@@ -1215,39 +1213,30 @@ static bool load_too_imbalanced(long src_load, long dst_load,
 	dst_capacity = env->dst_stats.compute_capacity;
 
 	/* We care about the slope of the imbalance, not the direction. */
-	load_a = dst_load;
-	load_b = src_load;
-	if (load_a < load_b)
-		swap(load_a, load_b);
+	if (dst_load < src_load)
+		swap(dst_load, src_load);
 
 	/* Is the difference below the threshold? */
-	imb = load_a * src_capacity * 100 -
-		load_b * dst_capacity * env->imbalance_pct;
+	imb = dst_load * src_capacity * 100 -
+	      src_load * dst_capacity * env->imbalance_pct;
 	if (imb <= 0)
 		return false;
 
 	/*
 	 * The imbalance is above the allowed threshold.
-	 * Allow a move that brings us closer to a balanced situation,
-	 * without moving things past the point of balance.
+	 * Compare it with the old imbalance.
 	 */
 	orig_src_load = env->src_stats.load;
+	orig_dst_load = env->dst_stats.load;
 
-	/*
-	 * In a task swap, there will be one load moving from src to dst,
-	 * and another moving back. This is the net sum of both moves.
-	 * A simple task move will always have a positive value.
-	 * Allow the move if it brings the system closer to a balanced
-	 * situation, without crossing over the balance point.
-	 */
-	moved_load = orig_src_load - src_load;
+	if (orig_dst_load < orig_src_load)
+		swap(orig_dst_load, orig_src_load);
 
-	if (moved_load > 0)
-		/* Moving src -> dst. Did we overshoot balance? */
-		return src_load * dst_capacity < dst_load * src_capacity;
-	else
-		/* Moving dst -> src. Did we overshoot balance? */
-		return dst_load * src_capacity < src_load * dst_capacity;
+	old_imb = orig_dst_load * src_capacity * 100 -
+		  orig_src_load * dst_capacity * env->imbalance_pct;
+
+	/* Would this change make things worse? */
+	return (imb > old_imb);
 }
 
 /*

2.1.0

From: Rik van Riel <[email protected]>

Changeset a43455a1 ("sched/numa: Ensure task_numa_migrate() checks the preferred node") fixes an issue where workloads would never converge on a fully loaded (or overloaded) system.

However, it introduces a regression on less than fully loaded systems, where workloads converge on a few NUMA nodes, instead of properly staying spread out across the whole system. This leads to a reduction in available memory bandwidth, and usable CPU cache, with predictable performance problems.

The root cause appears to be an interaction between the load balancer and NUMA balancing, where the short term load represented by the load balancer differs from the long term load the NUMA balancing code would like to base its decisions on.

Simply reverting a43455a1 would re-introduce the non-convergence of workloads on fully loaded systems, so that is not a good option. As an aside, the check done before a43455a1 only applied to a task's preferred node, not to other candidate nodes in the system, so the converge-on-too-few-nodes problem still happens, just to a lesser degree.

Instead, try to compensate for the impedance mismatch between the load balancer and NUMA balancing by only ever considering a lesser loaded node as a destination for NUMA balancing, regardless of whether the task is trying to move to the preferred node, or to another node.

This patch also addresses the issue that a system with a single runnable thread would never migrate that thread to near its memory, introduced by 095bebf61a46 ("sched/numa: Do not move past the balance point if unbalanced").

A test where the main thread creates a large memory area, and spawns a worker thread to iterate over the memory (placed on another node by select_task_rq_fair), after which the main thread goes to sleep and waits for the worker thread to loop over all the memory now sees the worker thread migrated to where the memory is, instead of having all the memory migrated over like before.

Jirka has run a number of performance tests on several systems:
single instance SpecJBB 2005 performance is 7-15% higher on a 4 node system, with higher gains on systems with more cores per socket. Multi-instance SpecJBB 2005 (one per node), linpack, and stream see little or no changes with the revert of 095bebf61a46 and this patch.

---
 kernel/sched/fair.c | 30 ++++++++++++++++++++++++++++--
 1 file changed, 28 insertions(+), 2 deletions(-)

index c47bf0dffb34..f655f2ad155d 100644
@@ -1398,6 +1398,30 @@ static void task_numa_find_cpu(struct task_numa_env *env,
 	}
 }
 
+/* Only move tasks to a NUMA node less busy than the current node. */
+static bool numa_has_capacity(struct task_numa_env *env)
+{
+	struct numa_stats *src = &env->src_stats;
+	struct numa_stats *dst = &env->dst_stats;
+
+	if (src->has_free_capacity && !dst->has_free_capacity)
+		return false;
+
+	/*
+	 * Only consider a task move if the source has a higher destination
+	 * than the destination, corrected for CPU capacity on each node.
+	 *
+	 *      src->load                dst->load
+	 * --------------------- vs ---------------------
+	 * src->compute_capacity    dst->compute_capacity
+	 */
+	if (src->load * dst->compute_capacity >
+	    dst->load * src->compute_capacity)
+		return true;
+
+	return false;
+}
+
 static int task_numa_migrate(struct task_struct *p)
 {
 	struct task_numa_env env = {

@@ -1452,7 +1476,8 @@ static int task_numa_migrate(struct task_struct *p)
 	update_numa_stats(&env.dst_stats, env.dst_nid);
 
 	/* Try to find a spot on the preferred nid. */
-	task_numa_find_cpu(&env, taskimp, groupimp);
+	if (numa_has_capacity(&env))
+		task_numa_find_cpu(&env, taskimp, groupimp);
 
 	/*
 	 * Look at other nodes in these cases:

@@ -1483,7 +1508,8 @@ static int task_numa_migrate(struct task_struct *p)
 			env.dist = dist;
 			env.dst_nid = nid;
 			update_numa_stats(&env.dst_stats, env.dst_nid);
-			task_numa_find_cpu(&env, taskimp, groupimp);
+			if (numa_has_capacity(&env))
+				task_numa_find_cpu(&env, taskimp, groupimp);
 		}
 	}
 

2.1.0

Uh yes, indeed. Thanks for spotting that.

Peter, do you want me to send a v2 with the typo fixed, or is it easier for you to fix up the typo before committing?

All rights reversed

Please send a new one, and also fix the Changelog to only include 12 char SHA1 references :-)

Here you are :)

---8<---

Subject: numa,sched: only consider less busy nodes as numa balancing destination

Changeset a43455a1d572 ("sched/numa: Ensure task_numa_migrate() checks the preferred node") fixes an issue where workloads would never converge on a fully loaded (or overloaded) system.

However, it introduces a regression on less than fully loaded systems, where workloads converge on a few NUMA nodes, instead of properly staying spread out across the whole system. This leads to a reduction in available memory bandwidth, and usable CPU cache, with predictable performance problems.

The root cause appears to be an interaction between the load balancer and NUMA balancing, where the short term load represented by the load balancer differs from the long term load the NUMA balancing code would like to base its decisions on.

Simply reverting a43455a1d572 would re-introduce the non-convergence of workloads on fully loaded systems, so that is not a good option. As an aside, the check done before a43455a1d572 only applied to a task's preferred node, not to other candidate nodes in the system, so the converge-on-too-few-nodes problem still happens, just to a lesser degree.

Instead, try to compensate for the impedance mismatch between the load balancer and NUMA balancing by only ever considering a lesser loaded node as a destination for NUMA balancing, regardless of whether the task is trying to move to the preferred node, or to another node.

This patch also addresses the issue that a system with a single runnable thread would never migrate that thread to near its memory, introduced by 095bebf61a46 ("sched/numa: Do not move past the balance point if unbalanced").

A test where the main thread creates a large memory area, and spawns a worker thread to iterate over the memory (placed on another node by select_task_rq_fair), after which the main thread goes to sleep and waits for the worker thread to loop over all the memory now sees the worker thread migrated to where the memory is, instead of having all the memory migrated over like before.

Jirka has run a number of performance tests on several systems:
single instance SpecJBB 2005 performance is 7-15% higher on a 4 node system, with higher gains on systems with more cores per socket. Multi-instance SpecJBB 2005 (one per node), linpack, and stream see little or no changes with the revert of 095bebf61a46 and this patch.

---
 kernel/sched/fair.c | 30 ++++++++++++++++++++++++++++--
 1 file changed, 28 insertions(+), 2 deletions(-)

index c47bf0dffb34..424d6ec1f61d 100644
@@ -1398,6 +1398,30 @@ static void task_numa_find_cpu(struct task_numa_env *env,
 	}
 }
 
+/* Only move tasks to a NUMA node less busy than the current node. */
+static bool numa_has_capacity(struct task_numa_env *env)
+{
+	struct numa_stats *src = &env->src_stats;
+	struct numa_stats *dst = &env->dst_stats;
+
+	if (src->has_free_capacity && !dst->has_free_capacity)
+		return false;
+
+	/*
+	 * Only consider a task move if the source has a higher load
+	 * than the destination, corrected for CPU capacity on each node.
+	 *
+	 *      src->load                dst->load
+	 * --------------------- vs ---------------------
+	 * src->compute_capacity    dst->compute_capacity
+	 */
+	if (src->load * dst->compute_capacity >
+	    dst->load * src->compute_capacity)
+		return true;
+
+	return false;
+}
+
 static int task_numa_migrate(struct task_struct *p)
 {
 	struct task_numa_env env = {

@@ -1452,7 +1476,8 @@ static int task_numa_migrate(struct task_struct *p)
 	update_numa_stats(&env.dst_stats, env.dst_nid);
 
 	/* Try to find a spot on the preferred nid. */
-	task_numa_find_cpu(&env, taskimp, groupimp);
+	if (numa_has_capacity(&env))
+		task_numa_find_cpu(&env, taskimp, groupimp);
 
 	/*
 	 * Look at other nodes in these cases:

@@ -1483,7 +1508,8 @@ static int task_numa_migrate(struct task_struct *p)
 			env.dist = dist;
 			env.dst_nid = nid;
 			update_numa_stats(&env.dst_stats, env.dst_nid);
-			task_numa_find_cpu(&env, taskimp, groupimp);
+			if (numa_has_capacity(&env))
+				task_numa_find_cpu(&env, taskimp, groupimp);
 		}
 	}
 

[Re-sending since it didn't hit the mailing list first time, due to HTML]

Hi Rik,

I've executed our eCommerce Web workload benchmark. Last time I did not revert 095bebf61a46, now I tested this patch-set. Let me summarize everything.

Here is the average web server response time in millisecs for various kernels.

1. v4.1-rc1 - 1450
2. v4.1-rc1 + a43455a1d572daf7b730fe12eb747d1e17411365 reverted - 300
3. v4.1-rc1 + NUMA disabled - 480
4. v4.1-rc5 + this patch-set - 1260

So as you see, for our workload reverting a43455a1d572daf7b730fe12eb747d1e17411365
results in Web server being most responsive (reminder - this is about a 2-socket Haswell-EP
server).

Just disabling NUMA also gives a big improvement, but not as good as reverting the
"offending" (from our workload's POW) patch.

This patch-set does result in a noticeable improvement too.

Artem.