From: Steve Sistare steven.sistare@oracle.com
maillist inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/I8PIYZ CVE: NA
Reference: https://lore.kernel.org/lkml/1541767840-93588-9-git-send-email-steven.sistar...
---------------------------
When a CPU has no more CFS tasks to run, and idle_balance() fails to find a task, then attempt to steal a task from an overloaded CPU in the same LLC, using the cfs_overload_cpus bitmap to efficiently identify candidates. To minimize search time, steal the first migratable task that is found when the bitmap is traversed. For fairness, search for migratable tasks on an overloaded CPU in order of next to run.
This simple stealing yields a higher CPU utilization than idle_balance() alone, because the search is cheap, so it may be called every time the CPU is about to go idle. idle_balance() does more work because it searches widely for the busiest queue, so to limit its CPU consumption, it declines to search if the system is too busy. Simple stealing does not offload the globally busiest queue, but it is much better than running nothing at all.
Stealing is controlled by the sched feature SCHED_STEAL, which is enabled by default.
Stealing imprroves utilization with only a modest CPU overhead in scheduler code. In the following experiment, hackbench is run with varying numbers of groups (40 tasks per group), and the delta in /proc/schedstat is shown for each run, averaged per CPU, augmented with these non-standard stats:
%find - percent of time spent in old and new functions that search for idle CPUs and tasks to steal and set the overloaded CPUs bitmap.
steal - number of times a task is stolen from another CPU.
X6-2: 1 socket * 10 cores * 2 hyperthreads = 20 CPUs Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz hackbench <grps> process 100000 sched_wakeup_granularity_ns=15000000
baseline grps time %busy slice sched idle wake %find steal 1 8.084 75.02 0.10 105476 46291 59183 0.31 0 2 13.892 85.33 0.10 190225 70958 119264 0.45 0 3 19.668 89.04 0.10 263896 87047 176850 0.49 0 4 25.279 91.28 0.10 322171 94691 227474 0.51 0 8 47.832 94.86 0.09 630636 144141 486322 0.56 0
new grps time %busy slice sched idle wake %find steal %speedup 1 5.938 96.80 0.24 31255 7190 24061 0.63 7433 36.1 2 11.491 99.23 0.16 74097 4578 69512 0.84 19463 20.9 3 16.987 99.66 0.15 115824 1985 113826 0.77 24707 15.8 4 22.504 99.80 0.14 167188 2385 164786 0.75 29353 12.3 8 44.441 99.86 0.11 389153 1616 387401 0.67 38190 7.6
Elapsed time improves by 8 to 36%, and CPU busy utilization is up by 5 to 22% hitting 99% for 2 or more groups (80 or more tasks). The cost is at most 0.4% more find time.
Additional performance results follow. A negative "speedup" is a regression. Note: for all hackbench runs, sched_wakeup_granularity_ns is set to 15 msec. Otherwise, preemptions increase at higher loads and distort the comparison between baseline and new.
------------------ 1 Socket Results ------------------
X6-2: 1 socket * 10 cores * 2 hyperthreads = 20 CPUs Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz Average of 10 runs of: hackbench <groups> process 100000
--- base -- --- new --- groups time %stdev time %stdev %speedup 1 8.008 0.1 5.905 0.2 35.6 2 13.814 0.2 11.438 0.1 20.7 3 19.488 0.2 16.919 0.1 15.1 4 25.059 0.1 22.409 0.1 11.8 8 47.478 0.1 44.221 0.1 7.3
X6-2: 1 socket * 22 cores * 2 hyperthreads = 44 CPUs Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz Average of 10 runs of: hackbench <groups> process 100000
--- base -- --- new --- groups time %stdev time %stdev %speedup 1 4.586 0.8 4.596 0.6 -0.3 2 7.693 0.2 5.775 1.3 33.2 3 10.442 0.3 8.288 0.3 25.9 4 13.087 0.2 11.057 0.1 18.3 8 24.145 0.2 22.076 0.3 9.3 16 43.779 0.1 41.741 0.2 4.8
KVM 4-cpu Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz tbench, average of 11 runs
clients %speedup 1 16.2 2 11.7 4 9.9 8 12.8 16 13.7
KVM 2-cpu Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz
Benchmark %speedup specjbb2015_critical_jops 5.7 mysql_sysb1.0.14_mutex_2 40.6 mysql_sysb1.0.14_oltp_2 3.9
------------------ 2 Socket Results ------------------
X6-2: 2 sockets * 10 cores * 2 hyperthreads = 40 CPUs Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz Average of 10 runs of: hackbench <groups> process 100000
--- base -- --- new --- groups time %stdev time %stdev %speedup 1 7.945 0.2 7.219 8.7 10.0 2 8.444 0.4 6.689 1.5 26.2 3 12.100 1.1 9.962 2.0 21.4 4 15.001 0.4 13.109 1.1 14.4 8 27.960 0.2 26.127 0.3 7.0
X6-2: 2 sockets * 22 cores * 2 hyperthreads = 88 CPUs Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz Average of 10 runs of: hackbench <groups> process 100000
--- base -- --- new --- groups time %stdev time %stdev %speedup 1 5.826 5.4 5.840 5.0 -0.3 2 5.041 5.3 6.171 23.4 -18.4 3 6.839 2.1 6.324 3.8 8.1 4 8.177 0.6 7.318 3.6 11.7 8 14.429 0.7 13.966 1.3 3.3 16 26.401 0.3 25.149 1.5 4.9
X6-2: 2 sockets * 22 cores * 2 hyperthreads = 88 CPUs Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz Oracle database OLTP, logging disabled, NVRAM storage
Customers Users %speedup 1200000 40 -1.2 2400000 80 2.7 3600000 120 8.9 4800000 160 4.4 6000000 200 3.0
X6-2: 2 sockets * 14 cores * 2 hyperthreads = 56 CPUs Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz Results from the Oracle "Performance PIT".
Benchmark %speedup
mysql_sysb1.0.14_fileio_56_rndrd 19.6 mysql_sysb1.0.14_fileio_56_seqrd 12.1 mysql_sysb1.0.14_fileio_56_rndwr 0.4 mysql_sysb1.0.14_fileio_56_seqrewr -0.3
pgsql_sysb1.0.14_fileio_56_rndrd 19.5 pgsql_sysb1.0.14_fileio_56_seqrd 8.6 pgsql_sysb1.0.14_fileio_56_rndwr 1.0 pgsql_sysb1.0.14_fileio_56_seqrewr 0.5
opatch_time_ASM_12.2.0.1.0_HP2M 7.5 select-1_users-warm_asmm_ASM_12.2.0.1.0_HP2M 5.1 select-1_users_asmm_ASM_12.2.0.1.0_HP2M 4.4 swingbenchv3_asmm_soebench_ASM_12.2.0.1.0_HP2M 5.8
lm3_memlat_L2 4.8 lm3_memlat_L1 0.0
ub_gcc_56CPUs-56copies_Pipe-based_Context_Switching 60.1 ub_gcc_56CPUs-56copies_Shell_Scripts_1_concurrent 5.2 ub_gcc_56CPUs-56copies_Shell_Scripts_8_concurrent -3.0 ub_gcc_56CPUs-56copies_File_Copy_1024_bufsize_2000_maxblocks 2.4
X5-2: 2 sockets * 18 cores * 2 hyperthreads = 72 CPUs Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz
NAS_OMP bench class ncpu %improved(Mops) dc B 72 1.3 is C 72 0.9 is D 72 0.7
sysbench mysql, average of 24 runs --- base --- --- new --- nthr events %stdev events %stdev %speedup 1 331.0 0.25 331.0 0.24 -0.1 2 661.3 0.22 661.8 0.22 0.0 4 1297.0 0.88 1300.5 0.82 0.2 8 2420.8 0.04 2420.5 0.04 -0.1 16 4826.3 0.07 4825.4 0.05 -0.1 32 8815.3 0.27 8830.2 0.18 0.1 64 12823.0 0.24 12823.6 0.26 0.0
-------------------------------------------------------------
Signed-off-by: Steve Sistare steven.sistare@oracle.com Signed-off-by: Cheng Yu serein.chengyu@huawei.com --- kernel/sched/fair.c | 165 ++++++++++++++++++++++++++++++++++++++-- kernel/sched/features.h | 6 ++ 2 files changed, 166 insertions(+), 5 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 7df713161c9f..4ea93402c41f 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -5075,6 +5075,9 @@ static void overload_clear(struct rq *rq) { struct sparsemask *overload_cpus;
+ if (!sched_feat(STEAL)) + return; + rcu_read_lock(); overload_cpus = rcu_dereference(rq->cfs_overload_cpus); if (overload_cpus) @@ -5086,6 +5089,9 @@ static void overload_set(struct rq *rq) { struct sparsemask *overload_cpus;
+ if (!sched_feat(STEAL)) + return; + rcu_read_lock(); overload_cpus = rcu_dereference(rq->cfs_overload_cpus); if (overload_cpus) @@ -5093,6 +5099,8 @@ static void overload_set(struct rq *rq) rcu_read_unlock(); }
+static int try_steal(struct rq *this_rq, struct rq_flags *rf); + #else /* CONFIG_SMP */
static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) @@ -5124,6 +5132,7 @@ static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
static inline void rq_idle_stamp_update(struct rq *rq) {} static inline void rq_idle_stamp_clear(struct rq *rq) {} +static inline int try_steal(struct rq *this_rq, struct rq_flags *rf) { return 0; } static inline void overload_clear(struct rq *rq) {} static inline void overload_set(struct rq *rq) {}
@@ -8792,21 +8801,23 @@ done: __maybe_unused; return NULL;
/* - * We must set idle_stamp _before_ calling idle_balance(), such that we - * measure the duration of idle_balance() as idle time. + * We must set idle_stamp _before_ calling try_steal() or + * idle_balance(), such that we measure the duration as idle time. */ rq_idle_stamp_update(rq);
new_tasks = newidle_balance(rq, rf); + if (new_tasks == 0) + new_tasks = try_steal(rq, rf);
if (new_tasks) rq_idle_stamp_clear(rq);
/* - * Because newidle_balance() releases (and re-acquires) rq->lock, it is - * possible for any higher priority task to appear. In that case we - * must re-start the pick_next_entity() loop. + * Because try_steal() and idle_balance() release (and re-acquire) + * rq->lock, it is possible for any higher priority task to appear. + * In that case we must re-start the pick_next_entity() loop. */ if (new_tasks < 0) return RETRY_TASK; @@ -12922,6 +12933,150 @@ static int task_is_throttled_fair(struct task_struct *p, int cpu) static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {} #endif
+/* + * Search the runnable tasks in @cfs_rq in order of next to run, and find + * the first one that can be migrated to @dst_rq. @cfs_rq is locked on entry. + * On success, dequeue the task from @cfs_rq and return it, else return NULL. + */ +static struct task_struct * +detach_next_task(struct cfs_rq *cfs_rq, struct rq *dst_rq) +{ + int dst_cpu = dst_rq->cpu; + struct task_struct *p; + struct rq *rq = rq_of(cfs_rq); + + lockdep_assert_rq_held(rq); + + list_for_each_entry_reverse(p, &rq->cfs_tasks, se.group_node) { + if (can_migrate_task_llc(p, rq, dst_rq)) { + detach_task(p, rq, dst_cpu); + return p; + } + } + return NULL; +} + +/* + * Attempt to migrate a CFS task from @src_cpu to @dst_rq. @locked indicates + * whether @dst_rq is already locked on entry. This function may lock or + * unlock @dst_rq, and updates @locked to indicate the locked state on return. + * The locking protocol is based on idle_balance(). + * Returns 1 on success and 0 on failure. + */ +static int steal_from(struct rq *dst_rq, struct rq_flags *dst_rf, bool *locked, + int src_cpu) +{ + struct task_struct *p; + struct rq_flags rf; + int stolen = 0; + int dst_cpu = dst_rq->cpu; + struct rq *src_rq = cpu_rq(src_cpu); + + if (dst_cpu == src_cpu || src_rq->cfs.h_nr_running < 2) + return 0; + + if (*locked) { + rq_unpin_lock(dst_rq, dst_rf); + raw_spin_rq_unlock(dst_rq); + *locked = false; + } + rq_lock_irqsave(src_rq, &rf); + update_rq_clock(src_rq); + + if (src_rq->cfs.h_nr_running < 2 || !cpu_active(src_cpu)) + p = NULL; + else + p = detach_next_task(&src_rq->cfs, dst_rq); + + rq_unlock(src_rq, &rf); + + if (p) { + raw_spin_rq_lock(dst_rq); + rq_repin_lock(dst_rq, dst_rf); + *locked = true; + update_rq_clock(dst_rq); + attach_task(dst_rq, p); + stolen = 1; + } + local_irq_restore(rf.flags); + + return stolen; +} + +/* + * Conservative upper bound on the max cost of a steal, in nsecs (the typical + * cost is 1-2 microsec). Do not steal if average idle time is less. + */ +#define SCHED_STEAL_COST 10000 + +/* + * Try to steal a runnable CFS task from a CPU in the same LLC as @dst_rq, + * and migrate it to @dst_rq. rq_lock is held on entry and return, but + * may be dropped in between. Return 1 on success, 0 on failure, and -1 + * if a task in a different scheduling class has become runnable on @dst_rq. + */ +static int try_steal(struct rq *dst_rq, struct rq_flags *dst_rf) +{ + int src_cpu; + int dst_cpu = dst_rq->cpu; + bool locked = true; + int stolen = 0; + struct sparsemask *overload_cpus; + + if (!sched_feat(STEAL)) + return 0; + + if (!cpu_active(dst_cpu)) + return 0; + + if (dst_rq->avg_idle < SCHED_STEAL_COST) + return 0; + + /* Get bitmap of overloaded CPUs in the same LLC as @dst_rq */ + + rcu_read_lock(); + overload_cpus = rcu_dereference(dst_rq->cfs_overload_cpus); + if (!overload_cpus) { + rcu_read_unlock(); + return 0; + } + +#ifdef CONFIG_SCHED_SMT + /* + * First try overloaded CPUs on the same core to preserve cache warmth. + */ + if (static_branch_likely(&sched_smt_present)) { + for_each_cpu(src_cpu, cpu_smt_mask(dst_cpu)) { + if (sparsemask_test_elem(overload_cpus, src_cpu) && + steal_from(dst_rq, dst_rf, &locked, src_cpu)) { + stolen = 1; + goto out; + } + } + } +#endif /* CONFIG_SCHED_SMT */ + + /* Accept any suitable task in the LLC */ + + sparsemask_for_each(overload_cpus, dst_cpu, src_cpu) { + if (steal_from(dst_rq, dst_rf, &locked, src_cpu)) { + stolen = 1; + goto out; + } + } + +out: + rcu_read_unlock(); + if (!locked) { + raw_spin_rq_lock(dst_rq); + rq_repin_lock(dst_rq, dst_rf); + } + stolen |= (dst_rq->cfs.h_nr_running > 0); + if (dst_rq->nr_running != dst_rq->cfs.h_nr_running) + stolen = -1; + return stolen; +} + /* * scheduler tick hitting a task of our scheduling class. * diff --git a/kernel/sched/features.h b/kernel/sched/features.h index f770168230ae..778afd56e466 100644 --- a/kernel/sched/features.h +++ b/kernel/sched/features.h @@ -52,6 +52,12 @@ SCHED_FEAT(TTWU_QUEUE, true) SCHED_FEAT(SIS_PROP, false) SCHED_FEAT(SIS_UTIL, true)
+/* + * Steal a CFS task from another CPU when going idle. + * Improves CPU utilization. + */ +SCHED_FEAT(STEAL, true) + /* * Issue a WARN when we do multiple update_rq_clock() calls * in a single rq->lock section. Default disabled because the