Skip to contentSkip to conversation

sched/fair: Scan cluster before scanning LLC in wake-up path

PATCH Superseded Yicong Yang 4 participants 10 messages Aug 22, 2022 – Sep 7, 2022 Scheduler

v7 of 11 Superseded Latest is v11 →

lore
Share & tools
lore
#0 lore raw

v7 cover sched/fair: Scan cluster before scanning LLC in wake-up path

From: Yicong Yang <[email protected]>

This is the follow-up work to support cluster scheduler. Previously we have added cluster level in the scheduler for both ARM64[1] and X86[2] to support load balance between clusters to bring more memory bandwidth and decrease cache contention. This patchset, on the other hand, takes care of wake-up path by giving CPUs within the same cluster a try before scanning the whole LLC to benefit those tasks communicating with each other.

[1] 778c558f49a2 ("sched: Add cluster scheduler level in core and related Kconfig for ARM64")
[2] 66558b730f25 ("sched: Add cluster scheduler level for x86")

Change since v6:
- rebase on 6.0-rc1

Change since v5:
- Improve patch 2 according to Peter's suggestion:
- use sched_cluster_active to indicate whether cluster is active
- consider SMT case and use wrap iteration when scanning cluster
- Add Vincent's tag
Thanks.

Change since v4:
- rename cpus_share_resources to cpus_share_lowest_cache to be more informative, per Tim
- return -1 when nr==0 in scan_cluster(), per Abel
Thanks!

Change since v3:
- fix compile error when !CONFIG_SCHED_CLUSTER, reported by lkp test.

Change since v2:
- leverage SIS_PROP to suspend redundant scanning when LLC is overloaded
- remove the ping-pong suppression
- address the comment from Tim, thanks.

Change since v1:
- regain the performance data based on v5.17-rc1
- rename cpus_share_cluster to cpus_share_resources per Vincent and Gautham, thanks!

Barry Song (2):
sched: Add per_cpu cluster domain info and cpus_share_lowest_cache API sched/fair: Scan cluster before scanning LLC in wake-up path

 include/linux/sched/sd_flags.h |  7 +++++++
 include/linux/sched/topology.h |  8 +++++++-
 kernel/sched/core.c            | 12 ++++++++++++
 kernel/sched/fair.c            | 30 +++++++++++++++++++++++++++---
 kernel/sched/sched.h           |  4 ++++
 kernel/sched/topology.c        | 25 +++++++++++++++++++++++++
 6 files changed, 82 insertions(+), 4 deletions(-)
Signature · 1 line
2.24.0
#1 lore raw

PATCH v7 1/2 sched: Add per_cpu cluster domain info and cpus_share_lowest_cache API

From: Barry Song <[email protected]>

Add per-cpu cluster domain info and cpus_share_lowest_cache() API. This is the preparation for the optimization of select_idle_cpu() on platforms with cluster scheduler level.

---
 include/linux/sched/sd_flags.h |  7 +++++++
 include/linux/sched/topology.h |  8 +++++++-
 kernel/sched/core.c            | 12 ++++++++++++
 kernel/sched/sched.h           |  2 ++
 kernel/sched/topology.c        | 15 +++++++++++++++
 5 files changed, 43 insertions(+), 1 deletion(-)
include/linux/sched/sd_flags.h +7 -0
index 57bde66d95f7..42ed454e8b18 100644
@@ -109,6 +109,13 @@ SD_FLAG(SD_ASYM_CPUCAPACITY_FULL, SDF_SHARED_PARENT | SDF_NEEDS_GROUPS)
  */
 SD_FLAG(SD_SHARE_CPUCAPACITY, SDF_SHARED_CHILD | SDF_NEEDS_GROUPS)
 
+/*
+ * Domain members share CPU cluster (LLC tags or L2 cache)
+ *
+ * NEEDS_GROUPS: Clusters are shared between groups.
+ */
+SD_FLAG(SD_CLUSTER, SDF_NEEDS_GROUPS)
+
 /*
  * Domain members share CPU package resources (i.e. caches)
  *
include/linux/sched/topology.h +7 -1
index 816df6cc444e..c0d21667ddf3 100644
@@ -45,7 +45,7 @@ static inline int cpu_smt_flags(void)
 #ifdef CONFIG_SCHED_CLUSTER
 static inline int cpu_cluster_flags(void)
 {
-	return SD_SHARE_PKG_RESOURCES;
+	return SD_CLUSTER | SD_SHARE_PKG_RESOURCES;
 }
 #endif
 
@@ -179,6 +179,7 @@ cpumask_var_t *alloc_sched_domains(unsigned int ndoms);
 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms);
 
 bool cpus_share_cache(int this_cpu, int that_cpu);
+bool cpus_share_lowest_cache(int this_cpu, int that_cpu);
 
 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
 typedef int (*sched_domain_flags_f)(void);
@@ -232,6 +233,11 @@ static inline bool cpus_share_cache(int this_cpu, int that_cpu)
 	return true;
 }
 
+static inline bool cpus_share_lowest_cache(int this_cpu, int that_cpu)
+{
+	return true;
+}
+
 #endif	/* !CONFIG_SMP */
 
 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
kernel/sched/core.c +12 -0
index ee28253c9ac0..2d647598d26c 100644
@@ -3802,6 +3802,18 @@ bool cpus_share_cache(int this_cpu, int that_cpu)
 	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
 }
 
+/*
+ * Whether CPUs are share lowest cache, which means LLC on non-cluster
+ * machines and LLC tag or L2 on machines with clusters.
+ */
+bool cpus_share_lowest_cache(int this_cpu, int that_cpu)
+{
+	if (this_cpu == that_cpu)
+		return true;
+
+	return per_cpu(sd_lowest_cache_id, this_cpu) == per_cpu(sd_lowest_cache_id, that_cpu);
+}
+
 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
 {
 	/*
kernel/sched/sched.h +2 -0
index e26688d387ae..e9f0935605e2 100644
@@ -1809,7 +1809,9 @@ static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 DECLARE_PER_CPU(int, sd_llc_size);
 DECLARE_PER_CPU(int, sd_llc_id);
+DECLARE_PER_CPU(int, sd_lowest_cache_id);
 DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
+DECLARE_PER_CPU(struct sched_domain __rcu *, sd_cluster);
 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
kernel/sched/topology.c +15 -0
index 8739c2a5a54e..8ab27c0d6d1f 100644
@@ -664,6 +664,8 @@ static void destroy_sched_domains(struct sched_domain *sd)
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
 DEFINE_PER_CPU(int, sd_llc_size);
 DEFINE_PER_CPU(int, sd_llc_id);
+DEFINE_PER_CPU(int, sd_lowest_cache_id);
+DEFINE_PER_CPU(struct sched_domain __rcu *, sd_cluster);
 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
@@ -689,6 +691,18 @@ static void update_top_cache_domain(int cpu)
 	per_cpu(sd_llc_id, cpu) = id;
 	rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
 
+	sd = lowest_flag_domain(cpu, SD_CLUSTER);
+	if (sd)
+		id = cpumask_first(sched_domain_span(sd));
+	rcu_assign_pointer(per_cpu(sd_cluster, cpu), sd);
+
+	/*
+	 * This assignment should be placed after the sd_llc_id as
+	 * we want this id equals to cluster id on cluster machines
+	 * but equals to LLC id on non-Cluster machines.
+	 */
+	per_cpu(sd_lowest_cache_id, cpu) = id;
+
 	sd = lowest_flag_domain(cpu, SD_NUMA);
 	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
 
@@ -1532,6 +1546,7 @@ static struct cpumask		***sched_domains_numa_masks;
  */
 #define TOPOLOGY_SD_FLAGS		\
 	(SD_SHARE_CPUCAPACITY	|	\
+	 SD_CLUSTER		|	\
 	 SD_SHARE_PKG_RESOURCES |	\
 	 SD_NUMA		|	\
 	 SD_ASYM_PACKING)
Signature · 1 line
2.24.0
#2 lore raw

PATCH v7 2/2 sched/fair: Scan cluster before scanning LLC in wake-up path

In reply to Yicong Yang

From: Barry Song <[email protected]>

For platforms having clusters like Kunpeng920, CPUs within the same cluster have lower latency when synchronizing and accessing shared resources like cache. Thus, this patch tries to find an idle cpu within the cluster of the target CPU before scanning the whole LLC to gain lower latency.

Testing has been done on Kunpeng920 by pinning tasks to one numa and two numa. On Kunpeng920, Each numa has 8 clusters and each cluster has 4 CPUs.

With this patch, We noticed enhancement on tbench within one numa or cross two numa.

On numa 0:

                             6.0-rc1                patched
Hmean     1        351.20 (   0.00%)      396.45 *  12.88%*
Hmean     2        700.43 (   0.00%)      793.76 *  13.32%*
Hmean     4       1404.42 (   0.00%)     1583.62 *  12.76%*
Hmean     8       2833.31 (   0.00%)     3147.85 *  11.10%*
Hmean     16      5501.90 (   0.00%)     6089.89 *  10.69%*
Hmean     32     10428.59 (   0.00%)    10619.63 *   1.83%*
Hmean     64      8223.39 (   0.00%)     8306.93 *   1.02%*
Hmean     128     7042.88 (   0.00%)     7068.03 *   0.36%*

On numa 0-1:
6.0-rc1 patched

Hmean     1        363.06 (   0.00%)      397.13 *   9.38%*
Hmean     2        721.68 (   0.00%)      789.84 *   9.44%*
Hmean     4       1435.15 (   0.00%)     1566.01 *   9.12%*
Hmean     8       2776.17 (   0.00%)     3007.05 *   8.32%*
Hmean     16      5471.71 (   0.00%)     6103.91 *  11.55%*
Hmean     32     10164.98 (   0.00%)    11531.81 *  13.45%*
Hmean     64     17143.28 (   0.00%)    20078.68 *  17.12%*
Hmean     128    14552.70 (   0.00%)    15156.41 *   4.15%*
Hmean     256    12827.37 (   0.00%)    13326.86 *   3.89%*

Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.

---
 kernel/sched/fair.c     | 30 +++++++++++++++++++++++++++---
 kernel/sched/sched.h    |  2 ++
 kernel/sched/topology.c | 10 ++++++++++
 3 files changed, 39 insertions(+), 3 deletions(-)
kernel/sched/fair.c +27 -3
index 914096c5b1ae..6fa77610d0f5 100644
@@ -6437,6 +6437,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
 		}
 	}
 
+	if (static_branch_unlikely(&sched_cluster_active)) {
+		struct sched_domain *sdc = rcu_dereference(per_cpu(sd_cluster, target));
+
+		if (sdc) {
+			for_each_cpu_wrap(cpu, sched_domain_span(sdc), target + 1) {
+				if (!cpumask_test_cpu(cpu, cpus))
+					continue;
+
+				if (has_idle_core) {
+					i = select_idle_core(p, cpu, cpus, &idle_cpu);
+					if ((unsigned int)i < nr_cpumask_bits)
+						return i;
+				} else {
+					if (--nr <= 0)
+						return -1;
+					idle_cpu = __select_idle_cpu(cpu, p);
+					if ((unsigned int)idle_cpu < nr_cpumask_bits)
+						return idle_cpu;
+				}
+			}
+			cpumask_andnot(cpus, cpus, sched_domain_span(sdc));
+		}
+	}
+
 	for_each_cpu_wrap(cpu, cpus, target + 1) {
 		if (has_idle_core) {
 			i = select_idle_core(p, cpu, cpus, &idle_cpu);
@@ -6444,7 +6468,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
 				return i;
 
 		} else {
-			if (!--nr)
+			if (--nr <= 0)
 				return -1;
 			idle_cpu = __select_idle_cpu(cpu, p);
 			if ((unsigned int)idle_cpu < nr_cpumask_bits)
@@ -6543,7 +6567,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
 	/*
 	 * If the previous CPU is cache affine and idle, don't be stupid:
 	 */
-	if (prev != target && cpus_share_cache(prev, target) &&
+	if (prev != target && cpus_share_lowest_cache(prev, target) &&
 	    (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
 	    asym_fits_capacity(task_util, prev))
 		return prev;
@@ -6569,7 +6593,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
 	p->recent_used_cpu = prev;
 	if (recent_used_cpu != prev &&
 	    recent_used_cpu != target &&
-	    cpus_share_cache(recent_used_cpu, target) &&
+	    cpus_share_lowest_cache(recent_used_cpu, target) &&
 	    (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
 	    cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
 	    asym_fits_capacity(task_util, recent_used_cpu)) {
kernel/sched/sched.h +2 -0
index e9f0935605e2..60e8a91e29d1 100644
@@ -1815,7 +1815,9 @@ DECLARE_PER_CPU(struct sched_domain __rcu *, sd_cluster);
 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
+
 extern struct static_key_false sched_asym_cpucapacity;
+extern struct static_key_false sched_cluster_active;
 
 struct sched_group_capacity {
 	atomic_t		ref;
kernel/sched/topology.c +10 -0
index 8ab27c0d6d1f..04ead3227201 100644
@@ -670,7 +670,9 @@ DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
+
 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
+DEFINE_STATIC_KEY_FALSE(sched_cluster_active);
 
 static void update_top_cache_domain(int cpu)
 {
@@ -2268,6 +2270,7 @@ build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *att
 	struct rq *rq = NULL;
 	int i, ret = -ENOMEM;
 	bool has_asym = false;
+	bool has_cluster = false;
 
 	if (WARN_ON(cpumask_empty(cpu_map)))
 		goto error;
@@ -2289,6 +2292,7 @@ build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *att
 			sd = build_sched_domain(tl, cpu_map, attr, sd, i);
 
 			has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
+			has_cluster |= sd->flags & SD_CLUSTER;
 
 			if (tl == sched_domain_topology)
 				*per_cpu_ptr(d.sd, i) = sd;
@@ -2399,6 +2403,9 @@ build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *att
 	if (has_asym)
 		static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
 
+	if (has_cluster)
+		static_branch_inc_cpuslocked(&sched_cluster_active);
+
 	if (rq && sched_debug_verbose) {
 		pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
 			cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
@@ -2498,6 +2505,9 @@ static void detach_destroy_domains(const struct cpumask *cpu_map)
 	if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
 		static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
 
+	if (rcu_access_pointer(per_cpu(sd_cluster, cpu)))
+		static_branch_dec_cpuslocked(&sched_cluster_active);
+
 	rcu_read_lock();
 	for_each_cpu(i, cpu_map)
 		cpu_attach_domain(NULL, &def_root_domain, i);
Signature · 1 line
2.24.0
Reviewed-by on v7 2/2
#3 lore raw
Quoted 56 lines · Yicong Yang

From: Barry Song <[email protected]>

For platforms having clusters like Kunpeng920, CPUs within the same cluster have lower latency when synchronizing and accessing shared resources like cache. Thus, this patch tries to find an idle cpu within the cluster of the target CPU before scanning the whole LLC to gain lower latency.

Testing has been done on Kunpeng920 by pinning tasks to one numa and two numa. On Kunpeng920, Each numa has 8 clusters and each cluster has 4 CPUs.

With this patch, We noticed enhancement on tbench within one numa or cross two numa.

On numa 0:

                             6.0-rc1                patched
Hmean     1        351.20 (   0.00%)      396.45 *  12.88%*
Hmean     2        700.43 (   0.00%)      793.76 *  13.32%*
Hmean     4       1404.42 (   0.00%)     1583.62 *  12.76%*
Hmean     8       2833.31 (   0.00%)     3147.85 *  11.10%*
Hmean     16      5501.90 (   0.00%)     6089.89 *  10.69%*
Hmean     32     10428.59 (   0.00%)    10619.63 *   1.83%*
Hmean     64      8223.39 (   0.00%)     8306.93 *   1.02%*
Hmean     128     7042.88 (   0.00%)     7068.03 *   0.36%*

On numa 0-1:
6.0-rc1 patched

Hmean     1        363.06 (   0.00%)      397.13 *   9.38%*
Hmean     2        721.68 (   0.00%)      789.84 *   9.44%*
Hmean     4       1435.15 (   0.00%)     1566.01 *   9.12%*
Hmean     8       2776.17 (   0.00%)     3007.05 *   8.32%*
Hmean     16      5471.71 (   0.00%)     6103.91 *  11.55%*
Hmean     32     10164.98 (   0.00%)    11531.81 *  13.45%*
Hmean     64     17143.28 (   0.00%)    20078.68 *  17.12%*
Hmean     128    14552.70 (   0.00%)    15156.41 *   4.15%*
Hmean     256    12827.37 (   0.00%)    13326.86 *   3.89%*

Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.

---
 kernel/sched/fair.c     | 30 +++++++++++++++++++++++++++---
 kernel/sched/sched.h    |  2 ++
 kernel/sched/topology.c | 10 ++++++++++
 3 files changed, 39 insertions(+), 3 deletions(-)
kernel/sched/fair.c +0 -0
index 914096c5b1ae..6fa77610d0f5 100644

kernel/sched/fair.c static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool · line 6444

Quoted 8 lines
kernel/sched/fair.c +5 -0
@@ -6437,6 +6437,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
 		}
 	}
 
+	if (static_branch_unlikely(&sched_cluster_active)) {
+		struct sched_domain *sdc = rcu_dereference(per_cpu(sd_cluster, target));
+
+		if (sdc) {
+			for_each_cpu_wrap(cpu, sched_domain_span(sdc), target + 1) {

Looks good to me. One minor question, why don't we use cpumask_and(cpus, sched_domain_span(sdc), cpus);

kernel/sched/fair.c static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool · line 6446

Quoted 2 lines
kernel/sched/fair.c +2 -0
+				if (!cpumask_test_cpu(cpu, cpus))
+					continue;

so above check can be removed in each loop? Besides may I know what version this patch is based on? since I failed to apply the patch on v6.0-rc2. Other than that:

thanks,
Chenyu

Quoted 16 lines
kernel/sched/fair.c +16 -0
+
+				if (has_idle_core) {
+					i = select_idle_core(p, cpu, cpus, &idle_cpu);
+					if ((unsigned int)i < nr_cpumask_bits)
+						return i;
+				} else {
+					if (--nr <= 0)
+						return -1;
+					idle_cpu = __select_idle_cpu(cpu, p);
+					if ((unsigned int)idle_cpu < nr_cpumask_bits)
+						return idle_cpu;
+				}
+			}
+			cpumask_andnot(cpus, cpus, sched_domain_span(sdc));
+		}
+	}
#4 lore raw
Quoted 67 lines · Chen Yu
Quoted 56 lines · Yicong Yang

From: Barry Song <[email protected]>

For platforms having clusters like Kunpeng920, CPUs within the same cluster have lower latency when synchronizing and accessing shared resources like cache. Thus, this patch tries to find an idle cpu within the cluster of the target CPU before scanning the whole LLC to gain lower latency.

Testing has been done on Kunpeng920 by pinning tasks to one numa and two numa. On Kunpeng920, Each numa has 8 clusters and each cluster has 4 CPUs.

With this patch, We noticed enhancement on tbench within one numa or cross two numa.

On numa 0:

                             6.0-rc1                patched
Hmean     1        351.20 (   0.00%)      396.45 *  12.88%*
Hmean     2        700.43 (   0.00%)      793.76 *  13.32%*
Hmean     4       1404.42 (   0.00%)     1583.62 *  12.76%*
Hmean     8       2833.31 (   0.00%)     3147.85 *  11.10%*
Hmean     16      5501.90 (   0.00%)     6089.89 *  10.69%*
Hmean     32     10428.59 (   0.00%)    10619.63 *   1.83%*
Hmean     64      8223.39 (   0.00%)     8306.93 *   1.02%*
Hmean     128     7042.88 (   0.00%)     7068.03 *   0.36%*

On numa 0-1:
6.0-rc1 patched

Hmean     1        363.06 (   0.00%)      397.13 *   9.38%*
Hmean     2        721.68 (   0.00%)      789.84 *   9.44%*
Hmean     4       1435.15 (   0.00%)     1566.01 *   9.12%*
Hmean     8       2776.17 (   0.00%)     3007.05 *   8.32%*
Hmean     16      5471.71 (   0.00%)     6103.91 *  11.55%*
Hmean     32     10164.98 (   0.00%)    11531.81 *  13.45%*
Hmean     64     17143.28 (   0.00%)    20078.68 *  17.12%*
Hmean     128    14552.70 (   0.00%)    15156.41 *   4.15%*
Hmean     256    12827.37 (   0.00%)    13326.86 *   3.89%*

Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.

---
 kernel/sched/fair.c     | 30 +++++++++++++++++++++++++++---
 kernel/sched/sched.h    |  2 ++
 kernel/sched/topology.c | 10 ++++++++++
 3 files changed, 39 insertions(+), 3 deletions(-)
kernel/sched/fair.c +0 -0
index 914096c5b1ae..6fa77610d0f5 100644

kernel/sched/fair.c static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool · line 6444

Quoted 8 lines
kernel/sched/fair.c +5 -0
@@ -6437,6 +6437,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
 		}
 	}
 
+	if (static_branch_unlikely(&sched_cluster_active)) {
+		struct sched_domain *sdc = rcu_dereference(per_cpu(sd_cluster, target));
+
+		if (sdc) {
+			for_each_cpu_wrap(cpu, sched_domain_span(sdc), target + 1) {

Looks good to me. One minor question, why don't we use cpumask_and(cpus, sched_domain_span(sdc), cpus);

Quoted 3 lines

kernel/sched/fair.c static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool · line 6446

Quoted 2 lines
kernel/sched/fair.c +2 -0
+				if (!cpumask_test_cpu(cpu, cpus))
+					continue;

so above check can be removed in each loop?

Since we'll need to recalculate the mask of rest CPUs to test in the LLC after scanning the cluster CPUs.

Quoted 3 lines

Besides may I know what version this patch
is based on? since I failed to apply the patch on v6.0-rc2. Other than that:

It's on 6.0-rc1 when sent but can be cleanly rebased on rc2:

yangyicong@ubuntu:~/mainline_linux/linux_sub_workspace$ git log --oneline -3
0079c27ba265 (HEAD -> topost-cls-v7, topost-cls-v6) sched/fair: Scan cluster before scanning LLC in wake-up path
1ecb9e322bd7 sched: Add per_cpu cluster domain info and cpus_share_lowest_cache API
1c23f9e627a7 (tag: v6.0-rc2, origin/master, origin/HEAD, master) Linux 6.0-rc2

So I'm not sure where's the problem...

Quoted 2 lines

Thanks!

Quoted 20 lines

thanks,
Chenyu

Quoted 6 lines
kernel/sched/fair.c +6 -0
+
+				if (has_idle_core) {
+					i = select_idle_core(p, cpu, cpus, &idle_cpu);
+					if ((unsigned int)i < nr_cpumask_bits)
+						return i;
+				} else {

kernel/sched/fair.c static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool · line 6462

Quoted 10 lines
kernel/sched/fair.c +10 -0
+					if (--nr <= 0)
+						return -1;
+					idle_cpu = __select_idle_cpu(cpu, p);
+					if ((unsigned int)idle_cpu < nr_cpumask_bits)
+						return idle_cpu;
+				}
+			}
+			cpumask_andnot(cpus, cpus, sched_domain_span(sdc));
+		}
+	}

.

#5 lore raw
Quoted 68 lines · Yicong Yang
Quoted 67 lines · Chen Yu
Quoted 56 lines · Yicong Yang

From: Barry Song <[email protected]>

For platforms having clusters like Kunpeng920, CPUs within the same cluster have lower latency when synchronizing and accessing shared resources like cache. Thus, this patch tries to find an idle cpu within the cluster of the target CPU before scanning the whole LLC to gain lower latency.

Testing has been done on Kunpeng920 by pinning tasks to one numa and two numa. On Kunpeng920, Each numa has 8 clusters and each cluster has 4 CPUs.

With this patch, We noticed enhancement on tbench within one numa or cross two numa.

On numa 0:

                             6.0-rc1                patched
Hmean     1        351.20 (   0.00%)      396.45 *  12.88%*
Hmean     2        700.43 (   0.00%)      793.76 *  13.32%*
Hmean     4       1404.42 (   0.00%)     1583.62 *  12.76%*
Hmean     8       2833.31 (   0.00%)     3147.85 *  11.10%*
Hmean     16      5501.90 (   0.00%)     6089.89 *  10.69%*
Hmean     32     10428.59 (   0.00%)    10619.63 *   1.83%*
Hmean     64      8223.39 (   0.00%)     8306.93 *   1.02%*
Hmean     128     7042.88 (   0.00%)     7068.03 *   0.36%*

On numa 0-1:
6.0-rc1 patched

Hmean     1        363.06 (   0.00%)      397.13 *   9.38%*
Hmean     2        721.68 (   0.00%)      789.84 *   9.44%*
Hmean     4       1435.15 (   0.00%)     1566.01 *   9.12%*
Hmean     8       2776.17 (   0.00%)     3007.05 *   8.32%*
Hmean     16      5471.71 (   0.00%)     6103.91 *  11.55%*
Hmean     32     10164.98 (   0.00%)    11531.81 *  13.45%*
Hmean     64     17143.28 (   0.00%)    20078.68 *  17.12%*
Hmean     128    14552.70 (   0.00%)    15156.41 *   4.15%*
Hmean     256    12827.37 (   0.00%)    13326.86 *   3.89%*

Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.

---
 kernel/sched/fair.c     | 30 +++++++++++++++++++++++++++---
 kernel/sched/sched.h    |  2 ++
 kernel/sched/topology.c | 10 ++++++++++
 3 files changed, 39 insertions(+), 3 deletions(-)
kernel/sched/fair.c +0 -0
index 914096c5b1ae..6fa77610d0f5 100644

kernel/sched/fair.c static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool · line 6444

Quoted 8 lines
kernel/sched/fair.c +5 -0
@@ -6437,6 +6437,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
 		}
 	}
 
+	if (static_branch_unlikely(&sched_cluster_active)) {
+		struct sched_domain *sdc = rcu_dereference(per_cpu(sd_cluster, target));
+
+		if (sdc) {
+			for_each_cpu_wrap(cpu, sched_domain_span(sdc), target + 1) {

Looks good to me. One minor question, why don't we use cpumask_and(cpus, sched_domain_span(sdc), cpus);

Quoted 6 lines
Quoted 3 lines

kernel/sched/fair.c static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool · line 6446

Quoted 2 lines
kernel/sched/fair.c +2 -0
+				if (!cpumask_test_cpu(cpu, cpus))
+					continue;

so above check can be removed in each loop?

Since we'll need to recalculate the mask of rest CPUs to test in the LLC after scanning the cluster CPUs.

I was thinking of introducing a temporary variable cpumask_and(cpus_cluster, sched_domain_span(sdc), cpus); and iterate this cpus_cluster in the loop. But since the cpus is reused, it is ok to be as it is.

Quoted 9 lines
Quoted 3 lines

Besides may I know what version this patch
is based on? since I failed to apply the patch on v6.0-rc2. Other than that:

It's on 6.0-rc1 when sent but can be cleanly rebased on rc2:

yangyicong@ubuntu:~/mainline_linux/linux_sub_workspace$ git log --oneline -3
0079c27ba265 (HEAD -> topost-cls-v7, topost-cls-v6) sched/fair: Scan cluster before scanning LLC in wake-up path
1ecb9e322bd7 sched: Add per_cpu cluster domain info and cpus_share_lowest_cache API

I did not apply 1/2, and that was why it failed I think. Thanks for explaination.

Thanks,
Chenyu

Quoted 29 lines

1c23f9e627a7 (tag: v6.0-rc2, origin/master, origin/HEAD, master) Linux 6.0-rc2

So I'm not sure where's the problem...

Quoted 2 lines

Thanks!

Quoted 20 lines

thanks,
Chenyu

Quoted 6 lines
kernel/sched/fair.c +6 -0
+
+				if (has_idle_core) {
+					i = select_idle_core(p, cpu, cpus, &idle_cpu);
+					if ((unsigned int)i < nr_cpumask_bits)
+						return i;
+				} else {

kernel/sched/fair.c static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool · line 6462

Quoted 10 lines
kernel/sched/fair.c +10 -0
+					if (--nr <= 0)
+						return -1;
+					idle_cpu = __select_idle_cpu(cpu, p);
+					if ((unsigned int)idle_cpu < nr_cpumask_bits)
+						return idle_cpu;
+				}
+			}
+			cpumask_andnot(cpus, cpus, sched_domain_span(sdc));
+		}
+	}

.

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#6 lore raw
In reply to Yicong Yang

a friendly ping...

Thanks.

Quoted 59 lines · Yicong Yang

From: Yicong Yang <[email protected]>

This is the follow-up work to support cluster scheduler. Previously we have added cluster level in the scheduler for both ARM64[1] and X86[2] to support load balance between clusters to bring more memory bandwidth and decrease cache contention. This patchset, on the other hand, takes care of wake-up path by giving CPUs within the same cluster a try before scanning the whole LLC to benefit those tasks communicating with each other.

[1] 778c558f49a2 ("sched: Add cluster scheduler level in core and related Kconfig for ARM64")
[2] 66558b730f25 ("sched: Add cluster scheduler level for x86")

Change since v6:
- rebase on 6.0-rc1

Change since v5:
- Improve patch 2 according to Peter's suggestion:
- use sched_cluster_active to indicate whether cluster is active
- consider SMT case and use wrap iteration when scanning cluster
- Add Vincent's tag
Thanks.

Change since v4:
- rename cpus_share_resources to cpus_share_lowest_cache to be more informative, per Tim
- return -1 when nr==0 in scan_cluster(), per Abel
Thanks!

Change since v3:
- fix compile error when !CONFIG_SCHED_CLUSTER, reported by lkp test.

Change since v2:
- leverage SIS_PROP to suspend redundant scanning when LLC is overloaded
- remove the ping-pong suppression
- address the comment from Tim, thanks.

Change since v1:
- regain the performance data based on v5.17-rc1
- rename cpus_share_cluster to cpus_share_resources per Vincent and Gautham, thanks!

Barry Song (2):
sched: Add per_cpu cluster domain info and cpus_share_lowest_cache API sched/fair: Scan cluster before scanning LLC in wake-up path

 include/linux/sched/sd_flags.h |  7 +++++++
 include/linux/sched/topology.h |  8 +++++++-
 kernel/sched/core.c            | 12 ++++++++++++
 kernel/sched/fair.c            | 30 +++++++++++++++++++++++++++---
 kernel/sched/sched.h           |  4 ++++
 kernel/sched/topology.c        | 25 +++++++++++++++++++++++++
 6 files changed, 82 insertions(+), 4 deletions(-)
reviewer Tested-by on the v7 series
#7 lore raw
In reply to Yicong Yang

Hello Yicong,

We've tested the series on a dual socket Zen3 system (2 x 64C/128T).

tl;dr

- The results look good and the changes do not affect the Zen3 machine which doesn't contain any sched domain with SD_CLUSTER flag set.

- With the latest BIOS, I don't see any regression due to the addition of the new per CPU variables.
We had observed a regression in tbench previously when testing the v4 of the series on the system with a slightly outdated BIOS
(https://lore.kernel.org/lkml/[email protected]/)
but that doesn't seem to be the case with the latest BIOS :)

Detailed results from the standard benchmarks are reported below.

Quoted 10 lines · Yicong Yang

From: Yicong Yang <[email protected]>

This is the follow-up work to support cluster scheduler. Previously we have added cluster level in the scheduler for both ARM64[1] and X86[2] to support load balance between clusters to bring more memory bandwidth and decrease cache contention. This patchset, on the other hand, takes care of wake-up path by giving CPUs within the same cluster a try before scanning the whole LLC to benefit those tasks communicating with each other.

Quoted 3 lines

[1] 778c558f49a2 ("sched: Add cluster scheduler level in core and related Kconfig for ARM64")
[2] 66558b730f25 ("sched: Add cluster scheduler level for x86")

Discussed below are the results from running standard benchmarks on a dual socket Zen3 (2 x 64C/128T) machine configured in different NPS modes.

NPS Modes are used to logically divide single socket into multiple NUMA region.
Following is the NUMA configuration for each NPS mode on the system:

NPS1: Each socket is a NUMA node.

    Total 2 NUMA nodes in the dual socket machine.

    Node 0: 0-63,   128-191
    Node 1: 64-127, 192-255

NPS2: Each socket is further logically divided into 2 NUMA regions.

    Total 4 NUMA nodes exist over 2 socket.
   
    Node 0: 0-31,   128-159
    Node 1: 32-63,  160-191
    Node 2: 64-95,  192-223
    Node 3: 96-127, 223-255

NPS4: Each socket is logically divided into 4 NUMA regions.

    Total 8 NUMA nodes exist over 2 socket.
   
    Node 0: 0-15,    128-143
    Node 1: 16-31,   144-159
    Node 2: 32-47,   160-175
    Node 3: 48-63,   176-191
    Node 4: 64-79,   192-207
    Node 5: 80-95,   208-223
    Node 6: 96-111,  223-231
    Node 7: 112-127, 232-255

Benchmark Results:

Kernel versions:
- tip: 5.19.0 tip sched/core
- cluster: 5.19.0 tip sched/core + both the patches of the series

When we started testing, the tip was at:
commit: 5531ecffa4b9 "sched: Add update_current_exec_runtime helper"

~~~~~~~~~~~~~
~ hackbench ~
~~~~~~~~~~~~~

NPS1

Test: tip cluster

 1-groups:	   4.31 (0.00 pct)	   4.31 (0.00 pct)
 2-groups:	   4.93 (0.00 pct)	   4.86 (1.41 pct)
 4-groups:	   5.38 (0.00 pct)	   5.36 (0.37 pct)
 8-groups:	   5.59 (0.00 pct)	   5.54 (0.89 pct)
16-groups:	   7.18 (0.00 pct)	   7.47 (-4.03 pct)

NPS2

Test: tip cluster

 1-groups:	   4.25 (0.00 pct)	   4.40 (-3.52 pct)
 2-groups:	   4.83 (0.00 pct)	   4.73 (2.07 pct)
 4-groups:	   5.25 (0.00 pct)	   5.18 (1.33 pct)
 8-groups:	   5.56 (0.00 pct)	   5.45 (1.97 pct)
16-groups:	   6.72 (0.00 pct)	   6.63 (1.33 pct)

NPS4

Test: tip cluster

 1-groups:	   4.24 (0.00 pct)	   4.23 (0.23 pct)
 2-groups:	   4.88 (0.00 pct)	   4.78 (2.04 pct)
 4-groups:	   5.30 (0.00 pct)	   5.25 (0.94 pct)
 8-groups:	   5.66 (0.00 pct)	   5.61 (0.88 pct)
16-groups:	   6.79 (0.00 pct)	   7.05 (-3.82 pct)

~~~~~~~~~~~~
~ schbench ~
~~~~~~~~~~~~

NPS1

#workers: tip cluster

  1:	  37.00 (0.00 pct)	     22.00 (40.54 pct)
  2:	  39.00 (0.00 pct)	     23.00 (41.02 pct)
  4:	  41.00 (0.00 pct)	     30.00 (26.82 pct)
  8:	  53.00 (0.00 pct)	     47.00 (11.32 pct)
 16:	  73.00 (0.00 pct)	     73.00 (0.00 pct)
 32:	 116.00 (0.00 pct)	    117.00 (-0.86 pct)
 64:	 217.00 (0.00 pct)	    221.00 (-1.84 pct)
128:	 477.00 (0.00 pct)	    444.00 (6.91 pct)

256: 1062.00 (0.00 pct) 1050.00 (1.12 pct)
512: 47552.00 (0.00 pct) 48576.00 (-2.15 pct)

NPS2

#workers: tip cluster

  1:	  20.00 (0.00 pct)	     20.00 (0.00 pct)
  2:	  22.00 (0.00 pct)	     23.00 (-4.54 pct)
  4:	  30.00 (0.00 pct)	     31.00 (-3.33 pct)
  8:	  46.00 (0.00 pct)	     49.00 (-6.52 pct)
 16:	  70.00 (0.00 pct)	     72.00 (-2.85 pct)
 32:	 120.00 (0.00 pct)	    118.00 (1.66 pct)
 64:	 215.00 (0.00 pct)	    216.00 (-0.46 pct)
128:	 482.00 (0.00 pct)	    449.00 (6.84 pct)
256:	1042.00 (0.00 pct)	    995.00 (4.51 pct)

512: 47552.00 (0.00 pct) 47296.00 (0.53 pct)

NPS4

#workers: tip cluster

  1:	  18.00 (0.00 pct)	     20.00 (-11.11 pct)
  2:	  23.00 (0.00 pct)	     22.00 (4.34 pct)
  4:	  27.00 (0.00 pct)	     30.00 (-11.11 pct)
  8:	  57.00 (0.00 pct)	     60.00 (-5.26 pct)
 16:	  76.00 (0.00 pct)	     84.00 (-10.52 pct)
 32:	 120.00 (0.00 pct)	    115.00 (4.16 pct)
 64:	 219.00 (0.00 pct)	    212.00 (3.19 pct)
128:	 459.00 (0.00 pct)	    442.00 (3.70 pct)
256:	1078.00 (0.00 pct)	    983.00 (8.81 pct)

512: 47040.00 (0.00 pct) 48192.00 (-2.44 pct)

Note: schbench displays lot of run to run variance for low worker count. This behavior is due to the timing of new-idle balance which is not consistent across runs.

~~~~~~~~~~
~ tbench ~
~~~~~~~~~~

NPS1

Clients: tip cluster

    1	   573.26 (0.00 pct)	   572.61 (-0.11 pct)
    2	  1131.19 (0.00 pct)	  1122.41 (-0.77 pct)
    4	  2100.07 (0.00 pct)	  2081.74 (-0.87 pct)
    8	  3809.88 (0.00 pct)	  3732.14 (-2.04 pct)

16 6560.72 (0.00 pct) 6289.22 (-4.13 pct)
32 12203.23 (0.00 pct) 11811.74 (-3.20 pct)
64 22389.81 (0.00 pct) 21587.79 (-3.58 pct)
128 32449.37 (0.00 pct) 32967.15 (1.59 pct)
256 58962.40 (0.00 pct) 56604.63 (-3.99 pct)
512 59608.71 (0.00 pct) 56529.95 (-5.16 pct) * (Machine Overloaded)
512 57925.05 (0.00 pct) 56697.38 (-2.11 pct) [Verification Run]
1024 58037.02 (0.00 pct) 55751.53 (-3.93 pct)

NPS2

Clients: tip cluster

    1	   574.20 (0.00 pct)	   572.49 (-0.29 pct)
    2	  1131.56 (0.00 pct)	  1149.53 (1.58 pct)
    4	  2132.26 (0.00 pct)	  2084.18 (-2.25 pct)
    8	  3812.20 (0.00 pct)	  3683.04 (-3.38 pct)

16 6457.61 (0.00 pct) 6340.70 (-1.81 pct)
32 12263.82 (0.00 pct) 11714.15 (-4.48 pct)
64 22224.11 (0.00 pct) 21226.34 (-4.48 pct)
128 33040.38 (0.00 pct) 32478.99 (-1.69 pct)
256 56547.25 (0.00 pct) 52915.71 (-6.42 pct) * (Machine Overloaded)
256 55631.80 (0.00 pct) 52905.99 (-4.89 pct) [Verification Run]
512 56220.67 (0.00 pct) 54735.69 (-2.64 pct)
1024 56048.88 (0.00 pct) 54426.63 (-2.89 pct)

NPS4

Clients: tip cluster

    1	   575.50 (0.00 pct)	   570.65 (-0.84 pct)
    2	  1138.70 (0.00 pct)	  1137.75 (-0.08 pct)
    4	  2070.66 (0.00 pct)	  2103.18 (1.57 pct)
    8	  3811.70 (0.00 pct)	  3573.52 (-6.24 pct) *
    8	  3769.53 (0.00 pct)      3653.05 (-3.09 pct) [Verification Run]

16 6312.80 (0.00 pct) 6212.41 (-1.59 pct)
32 11418.14 (0.00 pct) 11721.01 (2.65 pct)
64 19671.16 (0.00 pct) 20053.77 (1.94 pct)
128 30258.53 (0.00 pct) 32585.15 (7.68 pct)
256 55838.10 (0.00 pct) 51318.64 (-8.09 pct) * (Machine Overloaded)
256 54291.03 (0.00 pct) 54379.80 (0.16 pct) [Verification Run]
512 55586.44 (0.00 pct) 51538.93 (-7.28 pct) * (Machine Overloaded)
512 54190.04 (0.00 pct) 54096.16 (-0.17 pct) [Verification Run]
1024 56370.35 (0.00 pct) 50768.68 (-9.93 pct) * (Machine Overloaded)
1024 56498.36 (0.00 pct) 54661.85 (-3.25 pct) [Verification Run]

~~~~~~~~~~
~ stream ~
~~~~~~~~~~

NPS1

- 10 Runs:

Test: tip cluster
Copy: 332237.51 (0.00 pct) 338085.24 (1.76 pct)
Scale: 215236.94 (0.00 pct) 214179.72 (-0.49 pct)
Add: 250753.67 (0.00 pct) 251181.86 (0.17 pct)
Triad: 259467.60 (0.00 pct) 262541.92 (1.18 pct)

- 100 Runs:

Test: tip cluster
Copy: 329320.65 (0.00 pct) 336947.39 (2.31 pct)
Scale: 218102.78 (0.00 pct) 219617.85 (0.69 pct)
Add: 251283.30 (0.00 pct) 251918.03 (0.25 pct)
Triad: 258044.33 (0.00 pct) 261512.99 (1.34 pct)

NPS2

- 10 Runs:

Test: tip cluster
Copy: 336926.24 (0.00 pct) 324310.01 (-3.74 pct)
Scale: 220120.41 (0.00 pct) 212795.43 (-3.32 pct)
Add: 252428.34 (0.00 pct) 254355.80 (0.76 pct)
Triad: 274268.23 (0.00 pct) 261777.03 (-4.55 pct)

- 100 Runs:

Test:	      tip                  cluster
 Copy:   338126.49 (0.00 pct)    338947.03 (0.24 pct)
Scale:   230229.59 (0.00 pct)    229991.65 (-0.10 pct)
  Add:   253964.25 (0.00 pct)    264374.57 (4.09 pct)
Triad:   272176.19 (0.00 pct)    274587.35 (0.88 pct)

NPS4

- 10 Runs:

Test:	      tip                  cluster
 Copy:   367144.56 (0.00 pct)    375452.26 (2.26 pct)
Scale:   246928.04 (0.00 pct)    243651.53 (-1.32 pct)
  Add:   272096.30 (0.00 pct)    272845.33 (0.27 pct)
Triad:   286644.55 (0.00 pct)    290925.20 (1.49 pct)

- 100 Runs:

Test: tip cluster
Copy: 351980.15 (0.00 pct) 375854.72 (6.78 pct)
Scale: 254918.41 (0.00 pct) 255904.90 (0.38 pct)
Add: 272722.89 (0.00 pct) 274075.11 (0.49 pct)
Triad: 283340.94 (0.00 pct) 287608.77 (1.50 pct)

~~~~~~~~~~~~~~~~~~~~
~ Additional notes ~
~~~~~~~~~~~~~~~~~~~~

- schbench is know to have a noticeable run-to-run variation for lower worker counts and any improvements or regression observed can be safely ignored. The results are included to make sure there are no unnecessarily large regressions as a result of task pileup.

- tbench shows slight run to run variation with larger number of clients on both tip and patched kernel. This is expected as the machine is overloaded at that point (equivalent of two or more tasks per CPU). "Verification Run" shows none of these regressions are persistent.

Quoted 3 lines

[..snip..]

Overall, the changes look good and doesn't affect system without a SD_CLUSTER domain like the Zen3 system used during testing.

--
Thanks and Regards,
Prateek

#8 lore raw
Quoted 285 lines · K Prateek Nayak

Hello Yicong,

We've tested the series on a dual socket Zen3 system (2 x 64C/128T).

tl;dr

- The results look good and the changes do not affect the Zen3 machine which doesn't contain any sched domain with SD_CLUSTER flag set.

- With the latest BIOS, I don't see any regression due to the addition of the new per CPU variables.
We had observed a regression in tbench previously when testing the v4 of the series on the system with a slightly outdated BIOS
(https://lore.kernel.org/lkml/[email protected]/)
but that doesn't seem to be the case with the latest BIOS :)

Detailed results from the standard benchmarks are reported below.

Quoted 10 lines · Yicong Yang

From: Yicong Yang <[email protected]>

This is the follow-up work to support cluster scheduler. Previously we have added cluster level in the scheduler for both ARM64[1] and X86[2] to support load balance between clusters to bring more memory bandwidth and decrease cache contention. This patchset, on the other hand, takes care of wake-up path by giving CPUs within the same cluster a try before scanning the whole LLC to benefit those tasks communicating with each other.

Quoted 3 lines

[1] 778c558f49a2 ("sched: Add cluster scheduler level in core and related Kconfig for ARM64")
[2] 66558b730f25 ("sched: Add cluster scheduler level for x86")

Discussed below are the results from running standard benchmarks on a dual socket Zen3 (2 x 64C/128T) machine configured in different NPS modes.

NPS Modes are used to logically divide single socket into multiple NUMA region.
Following is the NUMA configuration for each NPS mode on the system:

NPS1: Each socket is a NUMA node.

    Total 2 NUMA nodes in the dual socket machine.

    Node 0: 0-63,   128-191
    Node 1: 64-127, 192-255

NPS2: Each socket is further logically divided into 2 NUMA regions.

    Total 4 NUMA nodes exist over 2 socket.
   
    Node 0: 0-31,   128-159
    Node 1: 32-63,  160-191
    Node 2: 64-95,  192-223
    Node 3: 96-127, 223-255

NPS4: Each socket is logically divided into 4 NUMA regions.

    Total 8 NUMA nodes exist over 2 socket.
   
    Node 0: 0-15,    128-143
    Node 1: 16-31,   144-159
    Node 2: 32-47,   160-175
    Node 3: 48-63,   176-191
    Node 4: 64-79,   192-207
    Node 5: 80-95,   208-223
    Node 6: 96-111,  223-231
    Node 7: 112-127, 232-255

Benchmark Results:

Kernel versions:
- tip: 5.19.0 tip sched/core
- cluster: 5.19.0 tip sched/core + both the patches of the series

When we started testing, the tip was at:
commit: 5531ecffa4b9 "sched: Add update_current_exec_runtime helper"

~~~~~~~~~~~~~
~ hackbench ~
~~~~~~~~~~~~~

NPS1

Test: tip cluster

 1-groups:	   4.31 (0.00 pct)	   4.31 (0.00 pct)
 2-groups:	   4.93 (0.00 pct)	   4.86 (1.41 pct)
 4-groups:	   5.38 (0.00 pct)	   5.36 (0.37 pct)
 8-groups:	   5.59 (0.00 pct)	   5.54 (0.89 pct)
16-groups:	   7.18 (0.00 pct)	   7.47 (-4.03 pct)

NPS2

Test: tip cluster

 1-groups:	   4.25 (0.00 pct)	   4.40 (-3.52 pct)
 2-groups:	   4.83 (0.00 pct)	   4.73 (2.07 pct)
 4-groups:	   5.25 (0.00 pct)	   5.18 (1.33 pct)
 8-groups:	   5.56 (0.00 pct)	   5.45 (1.97 pct)
16-groups:	   6.72 (0.00 pct)	   6.63 (1.33 pct)

NPS4

Test: tip cluster

 1-groups:	   4.24 (0.00 pct)	   4.23 (0.23 pct)
 2-groups:	   4.88 (0.00 pct)	   4.78 (2.04 pct)
 4-groups:	   5.30 (0.00 pct)	   5.25 (0.94 pct)
 8-groups:	   5.66 (0.00 pct)	   5.61 (0.88 pct)
16-groups:	   6.79 (0.00 pct)	   7.05 (-3.82 pct)

~~~~~~~~~~~~
~ schbench ~
~~~~~~~~~~~~

NPS1

#workers: tip cluster

  1:	  37.00 (0.00 pct)	     22.00 (40.54 pct)
  2:	  39.00 (0.00 pct)	     23.00 (41.02 pct)
  4:	  41.00 (0.00 pct)	     30.00 (26.82 pct)
  8:	  53.00 (0.00 pct)	     47.00 (11.32 pct)
 16:	  73.00 (0.00 pct)	     73.00 (0.00 pct)
 32:	 116.00 (0.00 pct)	    117.00 (-0.86 pct)
 64:	 217.00 (0.00 pct)	    221.00 (-1.84 pct)
128:	 477.00 (0.00 pct)	    444.00 (6.91 pct)

256: 1062.00 (0.00 pct) 1050.00 (1.12 pct)
512: 47552.00 (0.00 pct) 48576.00 (-2.15 pct)

NPS2

#workers: tip cluster

  1:	  20.00 (0.00 pct)	     20.00 (0.00 pct)
  2:	  22.00 (0.00 pct)	     23.00 (-4.54 pct)
  4:	  30.00 (0.00 pct)	     31.00 (-3.33 pct)
  8:	  46.00 (0.00 pct)	     49.00 (-6.52 pct)
 16:	  70.00 (0.00 pct)	     72.00 (-2.85 pct)
 32:	 120.00 (0.00 pct)	    118.00 (1.66 pct)
 64:	 215.00 (0.00 pct)	    216.00 (-0.46 pct)
128:	 482.00 (0.00 pct)	    449.00 (6.84 pct)
256:	1042.00 (0.00 pct)	    995.00 (4.51 pct)

512: 47552.00 (0.00 pct) 47296.00 (0.53 pct)

NPS4

#workers: tip cluster

  1:	  18.00 (0.00 pct)	     20.00 (-11.11 pct)
  2:	  23.00 (0.00 pct)	     22.00 (4.34 pct)
  4:	  27.00 (0.00 pct)	     30.00 (-11.11 pct)
  8:	  57.00 (0.00 pct)	     60.00 (-5.26 pct)
 16:	  76.00 (0.00 pct)	     84.00 (-10.52 pct)
 32:	 120.00 (0.00 pct)	    115.00 (4.16 pct)
 64:	 219.00 (0.00 pct)	    212.00 (3.19 pct)
128:	 459.00 (0.00 pct)	    442.00 (3.70 pct)
256:	1078.00 (0.00 pct)	    983.00 (8.81 pct)

512: 47040.00 (0.00 pct) 48192.00 (-2.44 pct)

Note: schbench displays lot of run to run variance for low worker count. This behavior is due to the timing of new-idle balance which is not consistent across runs.

~~~~~~~~~~
~ tbench ~
~~~~~~~~~~

NPS1

Clients: tip cluster

    1	   573.26 (0.00 pct)	   572.61 (-0.11 pct)
    2	  1131.19 (0.00 pct)	  1122.41 (-0.77 pct)
    4	  2100.07 (0.00 pct)	  2081.74 (-0.87 pct)
    8	  3809.88 (0.00 pct)	  3732.14 (-2.04 pct)

16 6560.72 (0.00 pct) 6289.22 (-4.13 pct)
32 12203.23 (0.00 pct) 11811.74 (-3.20 pct)
64 22389.81 (0.00 pct) 21587.79 (-3.58 pct)
128 32449.37 (0.00 pct) 32967.15 (1.59 pct)
256 58962.40 (0.00 pct) 56604.63 (-3.99 pct)
512 59608.71 (0.00 pct) 56529.95 (-5.16 pct) * (Machine Overloaded)
512 57925.05 (0.00 pct) 56697.38 (-2.11 pct) [Verification Run]
1024 58037.02 (0.00 pct) 55751.53 (-3.93 pct)

NPS2

Clients: tip cluster

    1	   574.20 (0.00 pct)	   572.49 (-0.29 pct)
    2	  1131.56 (0.00 pct)	  1149.53 (1.58 pct)
    4	  2132.26 (0.00 pct)	  2084.18 (-2.25 pct)
    8	  3812.20 (0.00 pct)	  3683.04 (-3.38 pct)

16 6457.61 (0.00 pct) 6340.70 (-1.81 pct)
32 12263.82 (0.00 pct) 11714.15 (-4.48 pct)
64 22224.11 (0.00 pct) 21226.34 (-4.48 pct)
128 33040.38 (0.00 pct) 32478.99 (-1.69 pct)
256 56547.25 (0.00 pct) 52915.71 (-6.42 pct) * (Machine Overloaded)
256 55631.80 (0.00 pct) 52905.99 (-4.89 pct) [Verification Run]
512 56220.67 (0.00 pct) 54735.69 (-2.64 pct)
1024 56048.88 (0.00 pct) 54426.63 (-2.89 pct)

NPS4

Clients: tip cluster

    1	   575.50 (0.00 pct)	   570.65 (-0.84 pct)
    2	  1138.70 (0.00 pct)	  1137.75 (-0.08 pct)
    4	  2070.66 (0.00 pct)	  2103.18 (1.57 pct)
    8	  3811.70 (0.00 pct)	  3573.52 (-6.24 pct) *
    8	  3769.53 (0.00 pct)      3653.05 (-3.09 pct) [Verification Run]

16 6312.80 (0.00 pct) 6212.41 (-1.59 pct)
32 11418.14 (0.00 pct) 11721.01 (2.65 pct)
64 19671.16 (0.00 pct) 20053.77 (1.94 pct)
128 30258.53 (0.00 pct) 32585.15 (7.68 pct)
256 55838.10 (0.00 pct) 51318.64 (-8.09 pct) * (Machine Overloaded)
256 54291.03 (0.00 pct) 54379.80 (0.16 pct) [Verification Run]
512 55586.44 (0.00 pct) 51538.93 (-7.28 pct) * (Machine Overloaded)
512 54190.04 (0.00 pct) 54096.16 (-0.17 pct) [Verification Run]
1024 56370.35 (0.00 pct) 50768.68 (-9.93 pct) * (Machine Overloaded)
1024 56498.36 (0.00 pct) 54661.85 (-3.25 pct) [Verification Run]

~~~~~~~~~~
~ stream ~
~~~~~~~~~~

NPS1

- 10 Runs:

Test: tip cluster
Copy: 332237.51 (0.00 pct) 338085.24 (1.76 pct)
Scale: 215236.94 (0.00 pct) 214179.72 (-0.49 pct)
Add: 250753.67 (0.00 pct) 251181.86 (0.17 pct)
Triad: 259467.60 (0.00 pct) 262541.92 (1.18 pct)

- 100 Runs:

Test: tip cluster
Copy: 329320.65 (0.00 pct) 336947.39 (2.31 pct)
Scale: 218102.78 (0.00 pct) 219617.85 (0.69 pct)
Add: 251283.30 (0.00 pct) 251918.03 (0.25 pct)
Triad: 258044.33 (0.00 pct) 261512.99 (1.34 pct)

NPS2

- 10 Runs:

Test: tip cluster
Copy: 336926.24 (0.00 pct) 324310.01 (-3.74 pct)
Scale: 220120.41 (0.00 pct) 212795.43 (-3.32 pct)
Add: 252428.34 (0.00 pct) 254355.80 (0.76 pct)
Triad: 274268.23 (0.00 pct) 261777.03 (-4.55 pct)

- 100 Runs:

Test:	      tip                  cluster
 Copy:   338126.49 (0.00 pct)    338947.03 (0.24 pct)
Scale:   230229.59 (0.00 pct)    229991.65 (-0.10 pct)
  Add:   253964.25 (0.00 pct)    264374.57 (4.09 pct)
Triad:   272176.19 (0.00 pct)    274587.35 (0.88 pct)

NPS4

- 10 Runs:

Test:	      tip                  cluster
 Copy:   367144.56 (0.00 pct)    375452.26 (2.26 pct)
Scale:   246928.04 (0.00 pct)    243651.53 (-1.32 pct)
  Add:   272096.30 (0.00 pct)    272845.33 (0.27 pct)
Triad:   286644.55 (0.00 pct)    290925.20 (1.49 pct)

- 100 Runs:

Test: tip cluster
Copy: 351980.15 (0.00 pct) 375854.72 (6.78 pct)
Scale: 254918.41 (0.00 pct) 255904.90 (0.38 pct)
Add: 272722.89 (0.00 pct) 274075.11 (0.49 pct)
Triad: 283340.94 (0.00 pct) 287608.77 (1.50 pct)

~~~~~~~~~~~~~~~~~~~~
~ Additional notes ~
~~~~~~~~~~~~~~~~~~~~

- schbench is know to have a noticeable run-to-run variation for lower worker counts and any improvements or regression observed can be safely ignored. The results are included to make sure there are no unnecessarily large regressions as a result of task pileup.

- tbench shows slight run to run variation with larger number of clients on both tip and patched kernel. This is expected as the machine is overloaded at that point (equivalent of two or more tasks per CPU). "Verification Run" shows none of these regressions are persistent.

Quoted 1 line
Quoted 8 lines
Quoted 2 lines

[..snip..]

Overall, the changes look good and doesn't affect system without a SD_CLUSTER domain like the Zen3 system used during testing.

Thanks a lot for the testing and verification on the Zen3 system.

Regards,
Yicong

#9 lore raw
In reply to Yicong Yang
Quoted 10 lines

From: Yicong Yang <[email protected]>

This is the follow-up work to support cluster scheduler. Previously we have added cluster level in the scheduler for both ARM64[1] and X86[2] to support load balance between clusters to bring more memory bandwidth and decrease cache contention. This patchset, on the other hand, takes care of wake-up path by giving CPUs within the same cluster a try before scanning the whole LLC to benefit those tasks communicating with each other.

Quoted 3 lines

Barry Song (2):
sched: Add per_cpu cluster domain info and cpus_share_lowest_cache API sched/fair: Scan cluster before scanning LLC in wake-up path

Hi Peter,
I believe this one has been ready. It has also gotten widely reviewed and tested on platforms w/ and w/o clusters.
Can you please pick up this?

Thanks
Barry