sched/fair: Scan cluster before scanning LLC in wake-up path
v9 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")
Since we're using sd->groups->flags to determine a cluster, core should ensure the flags set correctly on domain generation. This is done by [*].
[*] https://lore.kernel.org/all/[email protected]/
Change since v8:
- Peter find cpus_share_lowest_cache() is weired so fallback to cpus_share_resources() suggested in v4
- Use sd->groups->flags to find the cluster when scanning, save one per-cpu pointer
- Fix sched_cluster_active enabled incorrectly on domain degeneration
- Use sched_cluster_active to avoid repeated check on non-cluster machines, per Gautham
Change since v7:
- Optimize by choosing prev_cpu/recent_used_cpu when possible after failed to scanning for an idle CPU in cluster/LLC. Thanks Chen Yu for testing on Jacobsville
Change for RESEND:
- Collect tag from Chen Yu and rebase on the latest tip/sched/core. Thanks.
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 cpus_share_resources 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 | 59 +++++++++++++++++++++++++++++++--- kernel/sched/sched.h | 2 ++ kernel/sched/topology.c | 25 ++++++++++++++ 6 files changed, 107 insertions(+), 6 deletions(-)
Signature · 1 line
2.24.0
PATCH v9 1/2 sched: Add cpus_share_resources API
From: Barry Song <[email protected]>
Add cpus_share_resources() API. This is the preparation for the optimization of select_idle_cpu() on platforms with cluster scheduler level.
On a machine with clusters cpus_share_resources() will test whether two cpus are within the same cluster. On a non-cluster machine it will behaves the same as cpus_share_cache(). So we use "resources" here for cache resources.
--- include/linux/sched/sd_flags.h | 7 +++++++ include/linux/sched/topology.h | 8 +++++++- kernel/sched/core.c | 12 ++++++++++++ kernel/sched/sched.h | 1 + kernel/sched/topology.c | 13 +++++++++++++ 5 files changed, 40 insertions(+), 1 deletion(-)
include/linux/sched/sd_flags.h +7 -0
@@ -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
@@ -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_resources (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_resources (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
@@ -3953,6 +3953,18 @@ 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 cache resources, which means LLC on non-cluster +* machines and LLC tag or L2 on machines with clusters. +*/ +bool cpus_share_resources (int this_cpu,int that_cpu) +{ +if (this_cpu == that_cpu) +return true; + +return per_cpu (sd_share_id, this_cpu) ==per_cpu (sd_share_id, that_cpu); +} +static inline bool ttwu_queue_cond (struct task_struct *p,int cpu) {/*
kernel/sched/sched.h +1 -0
@@ -1847,6 +1847,7 @@ 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_share_id); DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared); DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa); DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
kernel/sched/topology.c +13 -0
@@ -666,6 +666,7 @@ 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_share_id); 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);
@@ -691,6 +692,17 @@ 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)); + +/* +* 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_share_id, cpu) = id; + sd =lowest_flag_domain (cpu,SD_NUMA );rcu_assign_pointer (per_cpu (sd_numa, cpu), sd);
@@ -1539,6 +1551,7 @@ 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
PATCH v9 2/2 sched/fair: Scan cluster before scanning LLC in wake-up path
In reply to Yicong YangFrom: 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. This will be implemented in 3 steps in select_idle_sibling():
1. When the prev_cpu/recent_used_cpu are good wakeup candidates, use them if they're sharing cluster with the target CPU. Otherwise record them and do the scanning first.
2. Scanning the cluster prior to the LLC of the target CPU for an idle CPU to wakeup.
3. If no idle CPU found after scanning and the prev_cpu/recent_used_cpu can be used, use them.
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 and netperf within one numa or cross two numa on 6.5-rc1:
tbench results (node 0):
baseline patched
1: 325.9673 378.9117 ( 16.24%)
4: 1311.9667 1501.5033 ( 14.45%)
8: 2629.4667 2961.9100 ( 12.64%)
16: 5259.1633 5928.0833 ( 12.72%)
32: 10368.6333 10566.8667 ( 1.91%)
64: 7868.7700 8182.0100 ( 3.98%)
128: 6528.5733 6801.8000 ( 4.19%)
tbench results (node 0-1):
vanilla patched
1: 329.2757 380.8907 ( 15.68%)
4: 1327.7900 1494.5300 ( 12.56%)
8: 2627.2133 2917.1233 ( 11.03%)
16: 5201.3367 5835.9233 ( 12.20%)
32: 8811.8500 11154.2000 ( 26.58%)
64: 15832.4000 19643.7667 ( 24.07%)
128: 12605.5667 14639.5667 ( 16.14%)
netperf results TCP_RR (node 0):
baseline patched
1: 77302.8667 92172.2100 ( 19.24%)
4: 78724.9200 91581.3100 ( 16.33%)
8: 79168.1296 91091.7942 ( 15.06%)
16: 81079.4200 90546.5225 ( 11.68%)
32: 82201.5799 78910.4982 ( -4.00%)
64: 29539.3509 29131.4698 ( -1.38%)
128: 12082.7522 11956.7705 ( -1.04%)
netperf results TCP_RR (node 0-1):
baseline patched
1: 78340.5233 92101.8733 ( 17.57%)
4: 79644.2483 91326.7517 ( 14.67%)
8: 79557.4313 90737.8096 ( 14.05%)
16: 79215.5304 90568.4542 ( 14.33%)
32: 78999.3983 85460.6044 ( 8.18%)
64: 74198.9494 74325.4361 ( 0.17%)
128: 27397.4810 27757.5471 ( 1.31%)
netperf results UDP_RR (node 0):
baseline patched
1: 95721.9367 111546.1367 ( 16.53%)
4: 96384.2250 110036.1408 ( 14.16%)
8: 97460.6546 109968.0883 ( 12.83%)
16: 98876.1687 109387.8065 ( 10.63%)
32: 104364.6417 105241.6767 ( 0.84%)
64: 37502.6246 37451.1204 ( -0.14%)
128: 14496.1780 14610.5538 ( 0.79%)
netperf results UDP_RR (node 0-1):
baseline patched
1: 96176.1633 111397.5333 ( 15.83%)
4: 94758.5575 105681.7833 ( 11.53%)
8: 94340.2200 104138.3613 ( 10.39%)
16: 95208.5285 106714.0396 ( 12.08%)
32: 74745.9028 100713.8764 ( 34.74%)
64: 59351.4977 73536.1434 ( 23.90%)
128: 23755.4971 26648.7413 ( 12.18%)Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.
Chen Yu also noticed this will improve the performance of tbench and netperf on a 24 CPUs Jacobsville machine, there are 4 CPUs in one cluster sharing L2 Cache.
--- kernel/sched/fair.c | 59 +++++++++++++++++++++++++++++++++++++---- kernel/sched/sched.h | 1 + kernel/sched/topology.c | 12 +++++++++ 3 files changed, 67 insertions(+), 5 deletions(-)
kernel/sched/fair.c +54 -5
@@ -7012,6 +7012,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool } } +if (static_branch_unlikely (&sched_cluster_active)) { +struct sched_group *sg = sd->groups; + +if (sg->flags &SD_CLUSTER ) { +for_each_cpu_wrap (cpu,sched_group_span (sg), 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_group_span (sg)); + } + } +for_each_cpu_wrap (cpu, cpus, target +1 ) {if (has_idle_core) { i =select_idle_core (p, cpu, cpus, &idle_cpu);
@@ -7019,7 +7043,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, boolreturn i; }else { -if (!--nr) +if (--nr <=0 )return -1 ; idle_cpu =__select_idle_cpu (cpu, p);if ((unsigned int )idle_cpu < nr_cpumask_bits)
@@ -7121,7 +7145,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)bool has_idle_core = false;struct sched_domain *sd;unsigned long task_util, util_min, util_max; -int i, recent_used_cpu; +int i, recent_used_cpu, prev_aff =-1 ;/* * On asymmetric system, update task utilization because we will check
@@ -7148,8 +7172,14 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)*/ if (prev != target &&cpus_share_cache (prev, target) && (available_idle_cpu (prev) ||sched_idle_cpu (prev)) && -asym_fits_cpu (task_util, util_min, util_max, prev)) -return prev; +asym_fits_cpu (task_util, util_min, util_max, prev)) { +if (!static_branch_unlikely (&sched_cluster_active)) +return prev; + +if (cpus_share_resources (prev, target)) +return prev; + prev_aff = prev; + }/* * Allow a per-cpu kthread to stack with the wakee if the
@@ -7176,7 +7206,13 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target) (available_idle_cpu (recent_used_cpu) ||sched_idle_cpu (recent_used_cpu)) &&cpumask_test_cpu (recent_used_cpu, p->cpus_ptr) && asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) { -return recent_used_cpu; + if (!static_branch_unlikely(&sched_cluster_active)) +return recent_used_cpu; + +if (cpus_share_resources (recent_used_cpu, target)) +return recent_used_cpu; + } else { + recent_used_cpu =-1 ; }/*
@@ -7217,6 +7253,19 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)if ((unsigned )i < nr_cpumask_bits)return i; +/* +* For cluster machines which have lower sharing cache like L2 or +* LLC Tag, we tend to find an idle CPU in the target's cluster +* first. But prev_cpu or recent_used_cpu may also be a good candidate, +* use them if possible when no idle CPU found in select_idle_cpu(). +*/ +if ((unsigned int )prev_aff < nr_cpumask_bits && + (available_idle_cpu (prev_aff) ||sched_idle_cpu (prev_aff))) +return prev_aff; +if ((unsigned int )recent_used_cpu < nr_cpumask_bits && + (available_idle_cpu (recent_used_cpu) ||sched_idle_cpu (recent_used_cpu))) +return recent_used_cpu; +return target; }
kernel/sched/sched.h +1 -0
@@ -1853,6 +1853,7 @@ 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; static __always_inline bool sched_asym_cpucap_active(void) {
kernel/sched/topology.c +12 -0
@@ -671,7 +671,9 @@ 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) {
@@ -2366,6 +2368,7 @@ build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attstruct rq *rq =NULL ;int i, ret = -ENOMEM ;bool has_asym = false; +bool has_cluster = false;if (WARN_ON (cpumask_empty (cpu_map))) goto error;
@@ -2491,12 +2494,18 @@ build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attWRITE_ONCE (d.rd->max_cpu_capacity, rq->cpu_capacity_orig);cpu_attach_domain (sd, d.rd, i); + +if (lowest_flag_domain (i,SD_CLUSTER )) + has_cluster = true; }rcu_read_unlock ();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);
@@ -2596,6 +2605,9 @@ 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 (static_branch_unlikely (&sched_cluster_active)) +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
Hi Yicong,
Thanks for sending this version!
Quoted 150 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. This will be implemented in 3 steps in
select_idle_sibling():
1. When the prev_cpu/recent_used_cpu are good wakeup candidates, use them if they're sharing cluster with the target CPU. Otherwise record them and do the scanning first.
2. Scanning the cluster prior to the LLC of the target CPU for an idle CPU to wakeup.
3. If no idle CPU found after scanning and the prev_cpu/recent_used_cpu can be used, use them.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 and netperf within one numa or cross two numa on 6.5-rc1:
tbench results (node 0):baseline patched 1: 325.9673 378.9117 ( 16.24%) 4: 1311.9667 1501.5033 ( 14.45%) 8: 2629.4667 2961.9100 ( 12.64%) 16: 5259.1633 5928.0833 ( 12.72%) 32: 10368.6333 10566.8667 ( 1.91%) 64: 7868.7700 8182.0100 ( 3.98%) 128: 6528.5733 6801.8000 ( 4.19%) tbench results (node 0-1): vanilla patched 1: 329.2757 380.8907 ( 15.68%) 4: 1327.7900 1494.5300 ( 12.56%) 8: 2627.2133 2917.1233 ( 11.03%) 16: 5201.3367 5835.9233 ( 12.20%) 32: 8811.8500 11154.2000 ( 26.58%) 64: 15832.4000 19643.7667 ( 24.07%) 128: 12605.5667 14639.5667 ( 16.14%) netperf results TCP_RR (node 0): baseline patched 1: 77302.8667 92172.2100 ( 19.24%) 4: 78724.9200 91581.3100 ( 16.33%) 8: 79168.1296 91091.7942 ( 15.06%) 16: 81079.4200 90546.5225 ( 11.68%) 32: 82201.5799 78910.4982 ( -4.00%) 64: 29539.3509 29131.4698 ( -1.38%) 128: 12082.7522 11956.7705 ( -1.04%) netperf results TCP_RR (node 0-1): baseline patched 1: 78340.5233 92101.8733 ( 17.57%) 4: 79644.2483 91326.7517 ( 14.67%) 8: 79557.4313 90737.8096 ( 14.05%) 16: 79215.5304 90568.4542 ( 14.33%) 32: 78999.3983 85460.6044 ( 8.18%) 64: 74198.9494 74325.4361 ( 0.17%) 128: 27397.4810 27757.5471 ( 1.31%) netperf results UDP_RR (node 0): baseline patched 1: 95721.9367 111546.1367 ( 16.53%) 4: 96384.2250 110036.1408 ( 14.16%) 8: 97460.6546 109968.0883 ( 12.83%) 16: 98876.1687 109387.8065 ( 10.63%) 32: 104364.6417 105241.6767 ( 0.84%) 64: 37502.6246 37451.1204 ( -0.14%) 128: 14496.1780 14610.5538 ( 0.79%) netperf results UDP_RR (node 0-1): baseline patched 1: 96176.1633 111397.5333 ( 15.83%) 4: 94758.5575 105681.7833 ( 11.53%) 8: 94340.2200 104138.3613 ( 10.39%) 16: 95208.5285 106714.0396 ( 12.08%) 32: 74745.9028 100713.8764 ( 34.74%) 64: 59351.4977 73536.1434 ( 23.90%) 128: 23755.4971 26648.7413 ( 12.18%)Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.
Chen Yu also noticed this will improve the performance of tbench and netperf on a 24 CPUs Jacobsville machine, there are 4 CPUs in one cluster sharing L2 Cache.
Suggested-by Peter Zijlstra <[email protected]>[https://lore.kernel.org/lkml/[email protected]]Tested-by Yicong Yang <[email protected]>Signed-off-by Barry Song <[email protected]>Signed-off-by Yicong Yang <[email protected]>Reviewed-by Tim Chen <[email protected]>Reviewed-by Chen Yu <[email protected]>--- kernel/sched/fair.c | 59 +++++++++++++++++++++++++++++++++++++---- kernel/sched/sched.h | 1 + kernel/sched/topology.c | 12 +++++++++ 3 files changed, 67 insertions(+), 5 deletions(-)kernel/sched/fair.c +26
-2@@ -7012,6 +7012,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool } } +if (static_branch_unlikely (&sched_cluster_active)) { +struct sched_group *sg = sd->groups; + +if (sg->flags &SD_CLUSTER ) { +for_each_cpu_wrap (cpu,sched_group_span (sg), 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_group_span (sg)); + } + } +for_each_cpu_wrap (cpu, cpus, target +1 ) {if (has_idle_core) { i =select_idle_core (p, cpu, cpus, &idle_cpu);@@ -7019,7 +7043,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, boolreturn i; }else { -if (!--nr) +if (--nr <=0 )return -1 ; idle_cpu =__select_idle_cpu (cpu, p);if ((unsigned int )idle_cpu < nr_cpumask_bits)@@ -7121,7 +7145,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)bool has_idle_core = false;struct sched_domain *sd;unsigned long task_util, util_min, util_max; -int i, recent_used_cpu; +int i, recent_used_cpu, prev_aff =-1 ;/* * On asymmetric system, update task utilization because we will check @@ -7148,8 +7172,14 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)*/
kernel/sched/fair.c static int select_idle_sibling(struct task_struct *p, int prev, int target) · line 7180
Quoted 10 lines
kernel/sched/fair.c +6
-2if (prev != target &&cpus_share_cache (prev, target) && (available_idle_cpu (prev) ||sched_idle_cpu (prev)) && -asym_fits_cpu (task_util, util_min, util_max, prev)) -return prev; +asym_fits_cpu (task_util, util_min, util_max, prev)) { +if (!static_branch_unlikely (&sched_cluster_active)) +return prev; + +if (cpus_share_resources (prev, target)) +return prev;
I have one minor question, previously Peter mentioned that he wants to get rid of the percpu sd_share_id, not sure if he means that not using it in select_idle_cpu() or remove that variable completely to not introduce extra space?
Hi Peter, could you please give us more hints on this? thanks.
If we wants to get rid of this variable, would this work?
if ((sd->groups->flags & SD_CLUSTER) && cpumask_test_cpu(prev, sched_group_span(sd->groups)) return prev
Anyway, I'll queue a test job on top of your version and see what the result is.
thanks,
Chenyu
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Hi Chenyu,
Sorry for the late reply. Something's wrong and cause this didn't appear in my mail box. I check it out on the LKML.
Quoted 171 lines · Chen Yu
Hi Yicong,
Thanks for sending this version!
Quoted 150 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. This will be implemented in 3 steps in
select_idle_sibling():
1. When the prev_cpu/recent_used_cpu are good wakeup candidates, use them if they're sharing cluster with the target CPU. Otherwise record them and do the scanning first.
2. Scanning the cluster prior to the LLC of the target CPU for an idle CPU to wakeup.
3. If no idle CPU found after scanning and the prev_cpu/recent_used_cpu can be used, use them.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 and netperf within one numa or cross two numa on 6.5-rc1:
tbench results (node 0):baseline patched 1: 325.9673 378.9117 ( 16.24%) 4: 1311.9667 1501.5033 ( 14.45%) 8: 2629.4667 2961.9100 ( 12.64%) 16: 5259.1633 5928.0833 ( 12.72%) 32: 10368.6333 10566.8667 ( 1.91%) 64: 7868.7700 8182.0100 ( 3.98%) 128: 6528.5733 6801.8000 ( 4.19%) tbench results (node 0-1): vanilla patched 1: 329.2757 380.8907 ( 15.68%) 4: 1327.7900 1494.5300 ( 12.56%) 8: 2627.2133 2917.1233 ( 11.03%) 16: 5201.3367 5835.9233 ( 12.20%) 32: 8811.8500 11154.2000 ( 26.58%) 64: 15832.4000 19643.7667 ( 24.07%) 128: 12605.5667 14639.5667 ( 16.14%) netperf results TCP_RR (node 0): baseline patched 1: 77302.8667 92172.2100 ( 19.24%) 4: 78724.9200 91581.3100 ( 16.33%) 8: 79168.1296 91091.7942 ( 15.06%) 16: 81079.4200 90546.5225 ( 11.68%) 32: 82201.5799 78910.4982 ( -4.00%) 64: 29539.3509 29131.4698 ( -1.38%) 128: 12082.7522 11956.7705 ( -1.04%) netperf results TCP_RR (node 0-1): baseline patched 1: 78340.5233 92101.8733 ( 17.57%) 4: 79644.2483 91326.7517 ( 14.67%) 8: 79557.4313 90737.8096 ( 14.05%) 16: 79215.5304 90568.4542 ( 14.33%) 32: 78999.3983 85460.6044 ( 8.18%) 64: 74198.9494 74325.4361 ( 0.17%) 128: 27397.4810 27757.5471 ( 1.31%) netperf results UDP_RR (node 0): baseline patched 1: 95721.9367 111546.1367 ( 16.53%) 4: 96384.2250 110036.1408 ( 14.16%) 8: 97460.6546 109968.0883 ( 12.83%) 16: 98876.1687 109387.8065 ( 10.63%) 32: 104364.6417 105241.6767 ( 0.84%) 64: 37502.6246 37451.1204 ( -0.14%) 128: 14496.1780 14610.5538 ( 0.79%) netperf results UDP_RR (node 0-1): baseline patched 1: 96176.1633 111397.5333 ( 15.83%) 4: 94758.5575 105681.7833 ( 11.53%) 8: 94340.2200 104138.3613 ( 10.39%) 16: 95208.5285 106714.0396 ( 12.08%) 32: 74745.9028 100713.8764 ( 34.74%) 64: 59351.4977 73536.1434 ( 23.90%) 128: 23755.4971 26648.7413 ( 12.18%)Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.
Chen Yu also noticed this will improve the performance of tbench and netperf on a 24 CPUs Jacobsville machine, there are 4 CPUs in one cluster sharing L2 Cache.
Suggested-by Peter Zijlstra <[email protected]>[https://lore.kernel.org/lkml/[email protected]]Tested-by Yicong Yang <[email protected]>Signed-off-by Barry Song <[email protected]>Signed-off-by Yicong Yang <[email protected]>Reviewed-by Tim Chen <[email protected]>Reviewed-by Chen Yu <[email protected]>--- kernel/sched/fair.c | 59 +++++++++++++++++++++++++++++++++++++---- kernel/sched/sched.h | 1 + kernel/sched/topology.c | 12 +++++++++ 3 files changed, 67 insertions(+), 5 deletions(-)kernel/sched/fair.c +26
-2@@ -7012,6 +7012,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool } } +if (static_branch_unlikely (&sched_cluster_active)) { +struct sched_group *sg = sd->groups; + +if (sg->flags &SD_CLUSTER ) { +for_each_cpu_wrap (cpu,sched_group_span (sg), 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_group_span (sg)); + } + } +for_each_cpu_wrap (cpu, cpus, target +1 ) {if (has_idle_core) { i =select_idle_core (p, cpu, cpus, &idle_cpu);@@ -7019,7 +7043,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, boolreturn i; }else { -if (!--nr) +if (--nr <=0 )return -1 ; idle_cpu =__select_idle_cpu (cpu, p);if ((unsigned int )idle_cpu < nr_cpumask_bits)@@ -7121,7 +7145,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)bool has_idle_core = false;struct sched_domain *sd;unsigned long task_util, util_min, util_max; -int i, recent_used_cpu; +int i, recent_used_cpu, prev_aff =-1 ;/* * On asymmetric system, update task utilization because we will check @@ -7148,8 +7172,14 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)*/ kernel/sched/fair.c static int select_idle_sibling(struct task_struct *p, int prev, int target) · line 7180
Quoted 10 lines
kernel/sched/fair.c +6
-2if (prev != target &&cpus_share_cache (prev, target) && (available_idle_cpu (prev) ||sched_idle_cpu (prev)) && -asym_fits_cpu (task_util, util_min, util_max, prev)) -return prev; +asym_fits_cpu (task_util, util_min, util_max, prev)) { +if (!static_branch_unlikely (&sched_cluster_active)) +return prev; + +if (cpus_share_resources (prev, target)) +return prev;I have one minor question, previously Peter mentioned that he wants to get rid of the percpu
sd_share_id, not sure if he means that not using it inselect_idle_cpu()or remove that variable completely to not introduce extra space?
Hi Peter, could you please give us more hints on this? thanks.
Quoted 6 lines
If we wants to get rid of this variable, would this work?
if ((sd->groups->flags & SD_CLUSTER) && cpumask_test_cpu(prev, sched_group_span(sd->groups)) return prev
In the current implementation, nop, we haven't deferenced the @sd yet and we don't need to if scanning is not needed.
Since we're on the quick path without scanning here, I wonder it'll be a bit more efficient to use a per-cpu id rather than deference the rcu and do the bitmap computation.
Thanks.
Chen Yu
Tested-byHi Yicong,
Quoted 5 lines · Yicong Yang
Hi Chenyu,
Sorry for the late reply. Something's wrong and cause this didn't appear in my mail box. I check it out on the LKML.
No worries : )
Quoted 179 lines · Chen Yu
Quoted 171 lines
Hi Yicong,
Thanks for sending this version!
Quoted 150 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. This will be implemented in 3 steps in
select_idle_sibling():
1. When the prev_cpu/recent_used_cpu are good wakeup candidates, use them if they're sharing cluster with the target CPU. Otherwise record them and do the scanning first.
2. Scanning the cluster prior to the LLC of the target CPU for an idle CPU to wakeup.
3. If no idle CPU found after scanning and the prev_cpu/recent_used_cpu can be used, use them.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 and netperf within one numa or cross two numa on 6.5-rc1:
tbench results (node 0):baseline patched 1: 325.9673 378.9117 ( 16.24%) 4: 1311.9667 1501.5033 ( 14.45%) 8: 2629.4667 2961.9100 ( 12.64%) 16: 5259.1633 5928.0833 ( 12.72%) 32: 10368.6333 10566.8667 ( 1.91%) 64: 7868.7700 8182.0100 ( 3.98%) 128: 6528.5733 6801.8000 ( 4.19%) tbench results (node 0-1): vanilla patched 1: 329.2757 380.8907 ( 15.68%) 4: 1327.7900 1494.5300 ( 12.56%) 8: 2627.2133 2917.1233 ( 11.03%) 16: 5201.3367 5835.9233 ( 12.20%) 32: 8811.8500 11154.2000 ( 26.58%) 64: 15832.4000 19643.7667 ( 24.07%) 128: 12605.5667 14639.5667 ( 16.14%) netperf results TCP_RR (node 0): baseline patched 1: 77302.8667 92172.2100 ( 19.24%) 4: 78724.9200 91581.3100 ( 16.33%) 8: 79168.1296 91091.7942 ( 15.06%) 16: 81079.4200 90546.5225 ( 11.68%) 32: 82201.5799 78910.4982 ( -4.00%) 64: 29539.3509 29131.4698 ( -1.38%) 128: 12082.7522 11956.7705 ( -1.04%) netperf results TCP_RR (node 0-1): baseline patched 1: 78340.5233 92101.8733 ( 17.57%) 4: 79644.2483 91326.7517 ( 14.67%) 8: 79557.4313 90737.8096 ( 14.05%) 16: 79215.5304 90568.4542 ( 14.33%) 32: 78999.3983 85460.6044 ( 8.18%) 64: 74198.9494 74325.4361 ( 0.17%) 128: 27397.4810 27757.5471 ( 1.31%) netperf results UDP_RR (node 0): baseline patched 1: 95721.9367 111546.1367 ( 16.53%) 4: 96384.2250 110036.1408 ( 14.16%) 8: 97460.6546 109968.0883 ( 12.83%) 16: 98876.1687 109387.8065 ( 10.63%) 32: 104364.6417 105241.6767 ( 0.84%) 64: 37502.6246 37451.1204 ( -0.14%) 128: 14496.1780 14610.5538 ( 0.79%) netperf results UDP_RR (node 0-1): baseline patched 1: 96176.1633 111397.5333 ( 15.83%) 4: 94758.5575 105681.7833 ( 11.53%) 8: 94340.2200 104138.3613 ( 10.39%) 16: 95208.5285 106714.0396 ( 12.08%) 32: 74745.9028 100713.8764 ( 34.74%) 64: 59351.4977 73536.1434 ( 23.90%) 128: 23755.4971 26648.7413 ( 12.18%)Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.
Chen Yu also noticed this will improve the performance of tbench and netperf on a 24 CPUs Jacobsville machine, there are 4 CPUs in one cluster sharing L2 Cache.
Suggested-by Peter Zijlstra <[email protected]>[https://lore.kernel.org/lkml/[email protected]]Tested-by Yicong Yang <[email protected]>Signed-off-by Barry Song <[email protected]>Signed-off-by Yicong Yang <[email protected]>Reviewed-by Tim Chen <[email protected]>Reviewed-by Chen Yu <[email protected]>--- kernel/sched/fair.c | 59 +++++++++++++++++++++++++++++++++++++---- kernel/sched/sched.h | 1 + kernel/sched/topology.c | 12 +++++++++ 3 files changed, 67 insertions(+), 5 deletions(-)kernel/sched/fair.c +26
-2@@ -7012,6 +7012,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool } } +if (static_branch_unlikely (&sched_cluster_active)) { +struct sched_group *sg = sd->groups; + +if (sg->flags &SD_CLUSTER ) { +for_each_cpu_wrap (cpu,sched_group_span (sg), 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_group_span (sg)); + } + } +for_each_cpu_wrap (cpu, cpus, target +1 ) {if (has_idle_core) { i =select_idle_core (p, cpu, cpus, &idle_cpu);@@ -7019,7 +7043,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, boolreturn i; }else { -if (!--nr) +if (--nr <=0 )return -1 ; idle_cpu =__select_idle_cpu (cpu, p);if ((unsigned int )idle_cpu < nr_cpumask_bits)@@ -7121,7 +7145,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)bool has_idle_core = false;struct sched_domain *sd;unsigned long task_util, util_min, util_max; -int i, recent_used_cpu; +int i, recent_used_cpu, prev_aff =-1 ;/* * On asymmetric system, update task utilization because we will check @@ -7148,8 +7172,14 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)*/ kernel/sched/fair.c static int select_idle_sibling(struct task_struct *p, int prev, int target) · line 7180
Quoted 10 lines
kernel/sched/fair.c +6
-2if (prev != target &&cpus_share_cache (prev, target) && (available_idle_cpu (prev) ||sched_idle_cpu (prev)) && -asym_fits_cpu (task_util, util_min, util_max, prev)) -return prev; +asym_fits_cpu (task_util, util_min, util_max, prev)) { +if (!static_branch_unlikely (&sched_cluster_active)) +return prev; + +if (cpus_share_resources (prev, target)) +return prev;I have one minor question, previously Peter mentioned that he wants to get rid of the percpu
sd_share_id, not sure if he means that not using it inselect_idle_cpu()or remove that variable completely to not introduce extra space?
Hi Peter, could you please give us more hints on this? thanks.Quoted 6 lines
If we wants to get rid of this variable, would this work?
if ((sd->groups->flags & SD_CLUSTER) && cpumask_test_cpu(prev, sched_group_span(sd->groups)) return prev
Quoted 7 lines
In the current implementation, nop, we haven't deferenced the @sd yet and we don't need to if scanning is not needed.
Since we're on the quick path without scanning here, I wonder it'll be a bit more efficient to use a per-cpu id rather than deference the rcu and do the bitmap computation.
Dereference is a memory barrier and the bitmap is of one operation/instruction which should not have too much overhead. But anyway I've tested this patch on Jacobsville and the data looks OK to me:
netperf
=======
case load baseline(std%) compare%( std%) TCP_RR 6-threads 1.00 ( 0.84) -0.32 ( 0.71) TCP_RR 12-threads 1.00 ( 0.35) +1.52 ( 0.42) TCP_RR 18-threads 1.00 ( 0.31) +3.89 ( 0.38) TCP_RR 24-threads 1.00 ( 0.87) -0.34 ( 0.75) TCP_RR 30-threads 1.00 ( 5.84) +0.71 ( 4.85) TCP_RR 36-threads 1.00 ( 4.84) +0.24 ( 3.30) TCP_RR 42-threads 1.00 ( 3.75) +0.26 ( 3.56) TCP_RR 48-threads 1.00 ( 1.51) +0.45 ( 1.28) UDP_RR 6-threads 1.00 ( 0.65) +10.12 ( 0.63) UDP_RR 12-threads 1.00 ( 0.20) +9.91 ( 0.25) UDP_RR 18-threads 1.00 ( 11.13) +16.77 ( 0.49) UDP_RR 24-threads 1.00 ( 12.38) +2.52 ( 0.98) UDP_RR 30-threads 1.00 ( 5.63) -0.34 ( 4.38) UDP_RR 36-threads 1.00 ( 19.12) -0.89 ( 3.30) UDP_RR 42-threads 1.00 ( 2.96) -1.41 ( 3.17) UDP_RR 48-threads 1.00 ( 14.08) -0.77 ( 10.77)
Good improvement in several cases. No regression is detected.
tbench
======
case load baseline(std%) compare%( std%) loopback 6-threads 1.00 ( 0.41) +1.63 ( 0.17) loopback 12-threads 1.00 ( 0.18) +4.39 ( 0.12) loopback 18-threads 1.00 ( 0.43) +10.42 ( 0.18) loopback 24-threads 1.00 ( 0.38) +1.24 ( 0.38) loopback 30-threads 1.00 ( 0.24) +0.60 ( 0.14) loopback 36-threads 1.00 ( 0.17) +0.63 ( 0.17) loopback 42-threads 1.00 ( 0.26) +0.76 ( 0.08) loopback 48-threads 1.00 ( 0.23) +0.91 ( 0.10)
Good improvement in 18-threads case. No regression is detected.
hackbench
=========
case load baseline(std%) compare%( std%) process-pipe 1-groups 1.00 ( 0.52) +9.26 ( 0.57) process-pipe 2-groups 1.00 ( 1.55) +6.92 ( 0.56) process-pipe 4-groups 1.00 ( 1.36) +4.80 ( 3.78) process-sockets 1-groups 1.00 ( 2.16) -6.35 ( 1.10) process-sockets 2-groups 1.00 ( 2.34) -6.35 ( 5.52) process-sockets 4-groups 1.00 ( 0.35) -5.64 ( 1.19) threads-pipe 1-groups 1.00 ( 0.82) +8.00 ( 0.00) threads-pipe 2-groups 1.00 ( 0.47) +6.91 ( 0.50) threads-pipe 4-groups 1.00 ( 0.45) +8.92 ( 2.27) threads-sockets 1-groups 1.00 ( 1.02) -4.13 ( 2.30) threads-sockets 2-groups 1.00 ( 0.34) -1.86 ( 2.39) threads-sockets 4-groups 1.00 ( 1.51) -3.99 ( 1.59)
Pros and cons for hackbench. There is improvement for pipe mode, but slight regression on sockets mode. I think this is within acceptable range.
schbench
========
case load baseline(std%) compare%( std%) normal 1-mthreads 1.00 ( 0.00) +0.00 ( 0.00) normal 2-mthreads 1.00 ( 0.00) +0.00 ( 0.00) normal 4-mthreads 1.00 ( 3.82) +0.00 ( 3.82)
There is impact to schbench at all, and the results are quite stable.
For the whole series:
thanks,
Chenyu
Quoted 265 lines · Chen Yu
Hi Yicong,
Quoted 5 lines · Yicong Yang
Hi Chenyu,
Sorry for the late reply. Something's wrong and cause this didn't appear in my mail box. I check it out on the LKML.
No worries : )
Quoted 179 lines · Chen Yu
Quoted 171 lines
Hi Yicong,
Thanks for sending this version!
Quoted 150 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. This will be implemented in 3 steps in
select_idle_sibling():
1. When the prev_cpu/recent_used_cpu are good wakeup candidates, use them if they're sharing cluster with the target CPU. Otherwise record them and do the scanning first.
2. Scanning the cluster prior to the LLC of the target CPU for an idle CPU to wakeup.
3. If no idle CPU found after scanning and the prev_cpu/recent_used_cpu can be used, use them.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 and netperf within one numa or cross two numa on 6.5-rc1:
tbench results (node 0):baseline patched 1: 325.9673 378.9117 ( 16.24%) 4: 1311.9667 1501.5033 ( 14.45%) 8: 2629.4667 2961.9100 ( 12.64%) 16: 5259.1633 5928.0833 ( 12.72%) 32: 10368.6333 10566.8667 ( 1.91%) 64: 7868.7700 8182.0100 ( 3.98%) 128: 6528.5733 6801.8000 ( 4.19%) tbench results (node 0-1): vanilla patched 1: 329.2757 380.8907 ( 15.68%) 4: 1327.7900 1494.5300 ( 12.56%) 8: 2627.2133 2917.1233 ( 11.03%) 16: 5201.3367 5835.9233 ( 12.20%) 32: 8811.8500 11154.2000 ( 26.58%) 64: 15832.4000 19643.7667 ( 24.07%) 128: 12605.5667 14639.5667 ( 16.14%) netperf results TCP_RR (node 0): baseline patched 1: 77302.8667 92172.2100 ( 19.24%) 4: 78724.9200 91581.3100 ( 16.33%) 8: 79168.1296 91091.7942 ( 15.06%) 16: 81079.4200 90546.5225 ( 11.68%) 32: 82201.5799 78910.4982 ( -4.00%) 64: 29539.3509 29131.4698 ( -1.38%) 128: 12082.7522 11956.7705 ( -1.04%) netperf results TCP_RR (node 0-1): baseline patched 1: 78340.5233 92101.8733 ( 17.57%) 4: 79644.2483 91326.7517 ( 14.67%) 8: 79557.4313 90737.8096 ( 14.05%) 16: 79215.5304 90568.4542 ( 14.33%) 32: 78999.3983 85460.6044 ( 8.18%) 64: 74198.9494 74325.4361 ( 0.17%) 128: 27397.4810 27757.5471 ( 1.31%) netperf results UDP_RR (node 0): baseline patched 1: 95721.9367 111546.1367 ( 16.53%) 4: 96384.2250 110036.1408 ( 14.16%) 8: 97460.6546 109968.0883 ( 12.83%) 16: 98876.1687 109387.8065 ( 10.63%) 32: 104364.6417 105241.6767 ( 0.84%) 64: 37502.6246 37451.1204 ( -0.14%) 128: 14496.1780 14610.5538 ( 0.79%) netperf results UDP_RR (node 0-1): baseline patched 1: 96176.1633 111397.5333 ( 15.83%) 4: 94758.5575 105681.7833 ( 11.53%) 8: 94340.2200 104138.3613 ( 10.39%) 16: 95208.5285 106714.0396 ( 12.08%) 32: 74745.9028 100713.8764 ( 34.74%) 64: 59351.4977 73536.1434 ( 23.90%) 128: 23755.4971 26648.7413 ( 12.18%)Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch in the code has not been tested but it supposed to work.
Chen Yu also noticed this will improve the performance of tbench and netperf on a 24 CPUs Jacobsville machine, there are 4 CPUs in one cluster sharing L2 Cache.
Suggested-by Peter Zijlstra <[email protected]>[https://lore.kernel.org/lkml/[email protected]]Tested-by Yicong Yang <[email protected]>Signed-off-by Barry Song <[email protected]>Signed-off-by Yicong Yang <[email protected]>Reviewed-by Tim Chen <[email protected]>Reviewed-by Chen Yu <[email protected]>--- kernel/sched/fair.c | 59 +++++++++++++++++++++++++++++++++++++---- kernel/sched/sched.h | 1 + kernel/sched/topology.c | 12 +++++++++ 3 files changed, 67 insertions(+), 5 deletions(-)kernel/sched/fair.c +26
-2@@ -7012,6 +7012,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool } } +if (static_branch_unlikely (&sched_cluster_active)) { +struct sched_group *sg = sd->groups; + +if (sg->flags &SD_CLUSTER ) { +for_each_cpu_wrap (cpu,sched_group_span (sg), 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_group_span (sg)); + } + } +for_each_cpu_wrap (cpu, cpus, target +1 ) {if (has_idle_core) { i =select_idle_core (p, cpu, cpus, &idle_cpu);@@ -7019,7 +7043,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, boolreturn i; }else { -if (!--nr) +if (--nr <=0 )return -1 ; idle_cpu =__select_idle_cpu (cpu, p);if ((unsigned int )idle_cpu < nr_cpumask_bits)@@ -7121,7 +7145,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)bool has_idle_core = false;struct sched_domain *sd;unsigned long task_util, util_min, util_max; -int i, recent_used_cpu; +int i, recent_used_cpu, prev_aff =-1 ;/* * On asymmetric system, update task utilization because we will check @@ -7148,8 +7172,14 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)*/ kernel/sched/fair.c static int select_idle_sibling(struct task_struct *p, int prev, int target) · line 7180
Quoted 10 lines
kernel/sched/fair.c +6
-2if (prev != target &&cpus_share_cache (prev, target) && (available_idle_cpu (prev) ||sched_idle_cpu (prev)) && -asym_fits_cpu (task_util, util_min, util_max, prev)) -return prev; +asym_fits_cpu (task_util, util_min, util_max, prev)) { +if (!static_branch_unlikely (&sched_cluster_active)) +return prev; + +if (cpus_share_resources (prev, target)) +return prev;I have one minor question, previously Peter mentioned that he wants to get rid of the percpu
sd_share_id, not sure if he means that not using it inselect_idle_cpu()or remove that variable completely to not introduce extra space?
Hi Peter, could you please give us more hints on this? thanks.Quoted 6 lines
If we wants to get rid of this variable, would this work?
if ((sd->groups->flags & SD_CLUSTER) && cpumask_test_cpu(prev, sched_group_span(sd->groups)) return prevQuoted 7 lines
In the current implementation, nop, we haven't deferenced the @sd yet and we don't need to if scanning is not needed.
Since we're on the quick path without scanning here, I wonder it'll be a bit more efficient to use a per-cpu id rather than deference the rcu and do the bitmap computation.
Dereference is a memory barrier and the bitmap is of one operation/instruction which should not have too much overhead. But anyway I've tested this patch on Jacobsville and the data looks OK to me:
netperf
=======case load baseline(std%) compare%( std%) TCP_RR 6-threads 1.00 ( 0.84) -0.32 ( 0.71) TCP_RR 12-threads 1.00 ( 0.35) +1.52 ( 0.42) TCP_RR 18-threads 1.00 ( 0.31) +3.89 ( 0.38) TCP_RR 24-threads 1.00 ( 0.87) -0.34 ( 0.75) TCP_RR 30-threads 1.00 ( 5.84) +0.71 ( 4.85) TCP_RR 36-threads 1.00 ( 4.84) +0.24 ( 3.30) TCP_RR 42-threads 1.00 ( 3.75) +0.26 ( 3.56) TCP_RR 48-threads 1.00 ( 1.51) +0.45 ( 1.28) UDP_RR 6-threads 1.00 ( 0.65) +10.12 ( 0.63) UDP_RR 12-threads 1.00 ( 0.20) +9.91 ( 0.25) UDP_RR 18-threads 1.00 ( 11.13) +16.77 ( 0.49) UDP_RR 24-threads 1.00 ( 12.38) +2.52 ( 0.98) UDP_RR 30-threads 1.00 ( 5.63) -0.34 ( 4.38) UDP_RR 36-threads 1.00 ( 19.12) -0.89 ( 3.30) UDP_RR 42-threads 1.00 ( 2.96) -1.41 ( 3.17) UDP_RR 48-threads 1.00 ( 14.08) -0.77 ( 10.77)Good improvement in several cases. No regression is detected.
tbench
======case load baseline(std%) compare%( std%) loopback 6-threads 1.00 ( 0.41) +1.63 ( 0.17) loopback 12-threads 1.00 ( 0.18) +4.39 ( 0.12) loopback 18-threads 1.00 ( 0.43) +10.42 ( 0.18) loopback 24-threads 1.00 ( 0.38) +1.24 ( 0.38) loopback 30-threads 1.00 ( 0.24) +0.60 ( 0.14) loopback 36-threads 1.00 ( 0.17) +0.63 ( 0.17) loopback 42-threads 1.00 ( 0.26) +0.76 ( 0.08) loopback 48-threads 1.00 ( 0.23) +0.91 ( 0.10)Good improvement in 18-threads case. No regression is detected.
hackbench
=========case load baseline(std%) compare%( std%) process-pipe 1-groups 1.00 ( 0.52) +9.26 ( 0.57) process-pipe 2-groups 1.00 ( 1.55) +6.92 ( 0.56) process-pipe 4-groups 1.00 ( 1.36) +4.80 ( 3.78) process-sockets 1-groups 1.00 ( 2.16) -6.35 ( 1.10) process-sockets 2-groups 1.00 ( 2.34) -6.35 ( 5.52) process-sockets 4-groups 1.00 ( 0.35) -5.64 ( 1.19) threads-pipe 1-groups 1.00 ( 0.82) +8.00 ( 0.00) threads-pipe 2-groups 1.00 ( 0.47) +6.91 ( 0.50) threads-pipe 4-groups 1.00 ( 0.45) +8.92 ( 2.27) threads-sockets 1-groups 1.00 ( 1.02) -4.13 ( 2.30) threads-sockets 2-groups 1.00 ( 0.34) -1.86 ( 2.39) threads-sockets 4-groups 1.00 ( 1.51) -3.99 ( 1.59)Pros and cons for hackbench. There is improvement for pipe mode, but slight regression on sockets mode. I think this is within acceptable range.
schbench
========case load baseline(std%) compare%( std%) normal 1-mthreads 1.00 ( 0.00) +0.00 ( 0.00) normal 2-mthreads 1.00 ( 0.00) +0.00 ( 0.00) normal 4-mthreads 1.00 ( 3.82) +0.00 ( 3.82)
Quoted 6 lines
There is impact to schbench at all, and the results are quite stable.
For the whole series:
Tested-by Chen Yu <[email protected]>
Thanks for testing.
Yicong.
Hi Peter,
A gentle ping for this. Any further comment?
Thanks.
Quoted 72 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 relatedKconfigfor ARM64")
[2]66558b730f25("sched: Add cluster scheduler level for x86")Since we're using
sd->groups->flagsto determine a cluster, core should ensure the flags set correctly on domain generation. This is done by [*].[*] https://lore.kernel.org/all/[email protected]/
Change since v8:
- Peter findcpus_share_lowest_cache()is weired so fallback tocpus_share_resources()suggested in v4
- Usesd->groups->flagsto find the cluster when scanning, save one per-cpu pointer
- Fixsched_cluster_activeenabled incorrectly on domain degeneration
- Usesched_cluster_activeto avoid repeated check on non-cluster machines, per GauthamChange since v7:
- Optimize by choosing prev_cpu/recent_used_cpu when possible after failed to scanning for an idle CPU in cluster/LLC. Thanks Chen Yu for testing on JacobsvilleChange for RESEND:
- Collect tag from Chen Yu and rebase on the latest tip/sched/core. Thanks.Change since v6:
- rebase on 6.0-rc1Change since v5:
- Improve patch 2 according to Peter's suggestion:
- usesched_cluster_activeto indicate whether cluster is active
- consider SMT case and use wrap iteration when scanning cluster
- Add Vincent's tag
Thanks.Change since v4:
- renamecpus_share_resourcestocpus_share_lowest_cacheto be more informative, per Tim
- return -1 when nr==0 inscan_cluster(), per Abel
Thanks!Change since v3:
- fix compile error when !CONFIG_SCHED_CLUSTER, reported by lkp test.Change since v2:
- leverageSIS_PROPto 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
- renamecpus_share_clustertocpus_share_resourcesper Vincent and Gautham, thanks!Barry Song (2):
sched: Addcpus_share_resourcesAPI
sched/fair: Scan cluster before scanning LLC in wake-up path
Quoted 8 lines
include/linux/sched/sd_flags.h | 7 ++++ include/linux/sched/topology.h | 8 ++++- kernel/sched/core.c | 12 +++++++ kernel/sched/fair.c | 59 +++++++++++++++++++++++++++++++--- kernel/sched/sched.h | 2 ++ kernel/sched/topology.c | 25 ++++++++++++++ 6 files changed, 107 insertions(+), 6 deletions(-)
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel