The bad thing about update_h_load(), which computes hierarchical load
factor for task groups, is that it is called for each task group in the
system before every load balancer run, and since rebalance can be
triggered very often, this function can eat really a lot of cpu time if
there are many cpu cgroups in the system.

Although the situation was improved significantly by commit a35b6466
('sched, cgroup: Reduce rq->lock hold times for large cgroup
hierarchies'), the problem still can arise under some kinds of loads,
e.g. when cpus are switching from idle to busy and back very frequently.

For instance, when I start 1000 of processes that wake up every
millisecond on my 8 cpus host, 'top' and 'perf top' show:

Cpu(s): 17.8%us, 24.3%sy,  0.0%ni, 57.9%id,  0.0%wa,  0.0%hi,  0.0%si
Events: 243K cycles
  7.57%  [kernel]               [k] __schedule
  7.08%  [kernel]               [k] timerqueue_add
  6.13%  libc-2.12.so           [.] usleep

Then if I create 10000 *idle* cpu cgroups (no processes in them), cpu
usage increases significantly although the 'wakers' are still executing
in the root cpu cgroup:

Cpu(s): 19.1%us, 48.7%sy,  0.0%ni, 31.6%id,  0.0%wa,  0.0%hi,  0.7%si
Events: 230K cycles
 24.56%  [kernel]            [k] tg_load_down
  5.76%  [kernel]            [k] __schedule

This happens because this particular kind of load triggers 'new idle'
rebalance very frequently, which requires calling update_h_load(),
which, in turn, calls tg_load_down() for every *idle* cpu cgroup even
though it is absolutely useless, because idle cpu cgroups have no tasks
to pull.

This patch tries to improve the situation by making h_load calculation
proceed only when h_load is really necessary. To achieve this, it
substitutes update_h_load() with update_cfs_rq_h_load(), which computes
h_load only for a given cfs_rq and all its ascendants, and makes the
load balancer call this function whenever it considers if a task should
be pulled, i.e. it moves h_load calculations directly to task_h_load().
For h_load of the same cfs_rq not to be updated multiple times (in case
several tasks in the same cgroup are considered during the same balance
run), the patch keeps the time of the last h_load update for each cfs_rq
and breaks calculation when it finds h_load to be uptodate.

The benefit of it is that h_load is computed only for those cfs_rq's,
which really need it, in particular all idle task groups are skipped.
Although this, in fact, moves h_load calculation under rq lock, it
should not affect latency much, because the amount of work done under rq
lock while trying to pull tasks is limited by sched_nr_migrate.

After the patch applied with the setup described above (1000 wakers in
the root cgroup and 10000 idle cgroups), I get:

Cpu(s): 16.9%us, 24.8%sy,  0.0%ni, 58.4%id,  0.0%wa,  0.0%hi,  0.0%si
Events: 242K cycles
  7.57%  [kernel]                  [k] __schedule
  6.70%  [kernel]                  [k] timerqueue_add
  5.93%  libc-2.12.so              [.] usleep
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.comSigned-off-by: Ingo Molnar <mingo@kernel.org>