sched/fair: Move the cache-hot 'load_avg' variable into its own cacheline

If a system with large number of sockets was driven to full
utilization, it was found that the clock tick handling occupied a
rather significant proportion of CPU time when fair group scheduling
and autogroup were enabled.

Running a java benchmark on a 16-socket IvyBridge-EX system, the perf
profile looked like:

  10.52%   0.00%  java   [kernel.vmlinux]  [k] smp_apic_timer_interrupt
   9.66%   0.05%  java   [kernel.vmlinux]  [k] hrtimer_interrupt
   8.65%   0.03%  java   [kernel.vmlinux]  [k] tick_sched_timer
   8.56%   0.00%  java   [kernel.vmlinux]  [k] update_process_times
   8.07%   0.03%  java   [kernel.vmlinux]  [k] scheduler_tick
   6.91%   1.78%  java   [kernel.vmlinux]  [k] task_tick_fair
   5.24%   5.04%  java   [kernel.vmlinux]  [k] update_cfs_shares

In particular, the high CPU time consumed by update_cfs_shares()
was mostly due to contention on the cacheline that contained the
task_group's load_avg statistical counter. This cacheline may also
contains variables like shares, cfs_rq & se which are accessed rather
frequently during clock tick processing.

This patch moves the load_avg variable into another cacheline
separated from the other frequently accessed variables. It also
creates a cacheline aligned kmemcache for task_group to make sure
that all the allocated task_group's are cacheline aligned.

By doing so, the perf profile became:

   9.44%   0.00%  java   [kernel.vmlinux]  [k] smp_apic_timer_interrupt
   8.74%   0.01%  java   [kernel.vmlinux]  [k] hrtimer_interrupt
   7.83%   0.03%  java   [kernel.vmlinux]  [k] tick_sched_timer
   7.74%   0.00%  java   [kernel.vmlinux]  [k] update_process_times
   7.27%   0.03%  java   [kernel.vmlinux]  [k] scheduler_tick
   5.94%   1.74%  java   [kernel.vmlinux]  [k] task_tick_fair
   4.15%   3.92%  java   [kernel.vmlinux]  [k] update_cfs_shares

The %cpu time is still pretty high, but it is better than before. The
benchmark results before and after the patch was as follows:

  Before patch - Max-jOPs: 907533    Critical-jOps: 134877
  After patch  - Max-jOPs: 916011    Critical-jOps: 142366

Signed-off-by: Waiman Long <Waiman.Long@hpe.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Ben Segall <bsegall@google.com>
Cc: Douglas Hatch <doug.hatch@hpe.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Paul Turner <pjt@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Scott J Norton <scott.norton@hpe.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Yuyang Du <yuyang.du@intel.com>
Link: http://lkml.kernel.org/r/1449081710-20185-3-git-send-email-Waiman.Long@hpe.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>

2 files changed, 13 insertions, 4 deletions

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index d591db1f2d6d..aa3f97869217 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -7370,6 +7370,9 @@ int in_sched_functions(unsigned long addr)
  */
 struct task_group root_task_group;
 LIST_HEAD(task_groups);
+
+/* Cacheline aligned slab cache for task_group */
+static struct kmem_cache *task_group_cache __read_mostly;
 #endif
 
 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
@@ -7427,11 +7430,12 @@ void __init sched_init(void)
 #endif /* CONFIG_RT_GROUP_SCHED */
 
 #ifdef CONFIG_CGROUP_SCHED
+	task_group_cache = KMEM_CACHE(task_group, 0);
+
 	list_add(&root_task_group.list, &task_groups);
 	INIT_LIST_HEAD(&root_task_group.children);
 	INIT_LIST_HEAD(&root_task_group.siblings);
 	autogroup_init(&init_task);
-
 #endif /* CONFIG_CGROUP_SCHED */
 
 	for_each_possible_cpu(i) {
@@ -7712,7 +7716,7 @@ static void free_sched_group(struct task_group *tg)
 	free_fair_sched_group(tg);
 	free_rt_sched_group(tg);
 	autogroup_free(tg);
-	kfree(tg);
+	kmem_cache_free(task_group_cache, tg);
 }
 
 /* allocate runqueue etc for a new task group */
@@ -7720,7 +7724,7 @@ struct task_group *sched_create_group(struct task_group *parent)
 {
 	struct task_group *tg;
 
-	tg = kzalloc(sizeof(*tg), GFP_KERNEL);
+	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
 	if (!tg)
 		return ERR_PTR(-ENOMEM);
 
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 472cd142e4f4..a5a6b3e60868 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -248,7 +248,12 @@ struct task_group {
 	unsigned long shares;
 
 #ifdef	CONFIG_SMP
-	atomic_long_t load_avg;
+	/*
+	 * load_avg can be heavily contended at clock tick time, so put
+	 * it in its own cacheline separated from the fields above which
+	 * will also be accessed at each tick.
+	 */
+	atomic_long_t load_avg ____cacheline_aligned;
 #endif
 #endif