Skip to content

L
linux
Switch branch/tag
  1. 07 Aug, 2013 29 commits
  3. 30 Jul, 2013 8 commits
  5. 23 Jul, 2013 3 commits
    • Peter Zijlstra's avatar
      sched: Micro-optimize the smart wake-affine logic · 7d9ffa89
      Peter Zijlstra authored 
      Smart wake-affine is using node-size as the factor currently, but the overhead
of the mask operation is high.

Thus, this patch introduce the 'sd_llc_size' percpu variable, which will record
the highest cache-share domain size, and make it to be the new factor, in order
to reduce the overhead and make it more reasonable.
Tested-by: default avatarDavidlohr Bueso <davidlohr.bueso@hp.com>
Tested-by: default avatarMichael Wang <wangyun@linux.vnet.ibm.com>
Signed-off-by: default avatarPeter Zijlstra <peterz@infradead.org>
Acked-by: default avatarMichael Wang <wangyun@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Link: http://lkml.kernel.org/r/51D5008E.6030102@linux.vnet.ibm.com
[ Tidied up the changelog. ]
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      7d9ffa89
    • Michael Wang's avatar
      sched: Implement smarter wake-affine logic · 62470419
      Michael Wang authored 
      The wake-affine scheduler feature is currently always trying to pull
the wakee close to the waker. In theory this should be beneficial if
the waker's CPU caches hot data for the wakee, and it's also beneficial
in the extreme ping-pong high context switch rate case.

Testing shows it can benefit hackbench up to 15%.

However, the feature is somewhat blind, from which some workloads
such as pgbench suffer. It's also time-consuming algorithmically.

Testing shows it can damage pgbench up to 50% - far more than the
benefit it brings in the best case.

So wake-affine should be smarter and it should realize when to
stop its thankless effort at trying to find a suitable CPU to wake on.

This patch introduces 'wakee_flips', which will be increased each
time the task flips (switches) its wakee target.

So a high 'wakee_flips' value means the task has more than one
wakee, and the bigger the number, the higher the wakeup frequency.

Now when making the decision on whether to pull or not, pay attention to
the wakee with a high 'wakee_flips', pulling such a task may benefit
the wakee. Also imply that the waker will face cruel competition later,
it could be very cruel or very fast depends on the story behind
'wakee_flips', waker therefore suffers.

Furthermore, if waker also has a high 'wakee_flips', that implies that
multiple tasks rely on it, then waker's higher latency will damage all
of them, so pulling wakee seems to be a bad deal.

Thus, when 'waker->wakee_flips / wakee->wakee_flips' becomes
higher and higher, the cost of pulling seems to be worse and worse.

The patch therefore helps the wake-affine feature to stop its pulling
work when:

	wakee->wakee_flips > factor &&
	waker->wakee_flips > (factor * wakee->wakee_flips)

The 'factor' here is the number of CPUs in the current CPU's NUMA node,
so a bigger node will lead to more pulling since the trial becomes more
severe.

After applying the patch, pgbench shows up to 40% improvements and no regressions.

Tested with 12 cpu x86 server and tip 3.10.0-rc7.

The percentages in the final column highlight the areas with the biggest wins,
all other areas improved as well:

	pgbench		    base	smart

	| db_size | clients |  tps  |	|  tps  |
	+---------+---------+-------+   +-------+
	| 22 MB   |       1 | 10598 |   | 10796 |
	| 22 MB   |       2 | 21257 |   | 21336 |
	| 22 MB   |       4 | 41386 |   | 41622 |
	| 22 MB   |       8 | 51253 |   | 57932 |
	| 22 MB   |      12 | 48570 |   | 54000 |
	| 22 MB   |      16 | 46748 |   | 55982 | +19.75%
	| 22 MB   |      24 | 44346 |   | 55847 | +25.93%
	| 22 MB   |      32 | 43460 |   | 54614 | +25.66%
	| 7484 MB |       1 |  8951 |   |  9193 |
	| 7484 MB |       2 | 19233 |   | 19240 |
	| 7484 MB |       4 | 37239 |   | 37302 |
	| 7484 MB |       8 | 46087 |   | 50018 |
	| 7484 MB |      12 | 42054 |   | 48763 |
	| 7484 MB |      16 | 40765 |   | 51633 | +26.66%
	| 7484 MB |      24 | 37651 |   | 52377 | +39.11%
	| 7484 MB |      32 | 37056 |   | 51108 | +37.92%
	| 15 GB   |       1 |  8845 |   |  9104 |
	| 15 GB   |       2 | 19094 |   | 19162 |
	| 15 GB   |       4 | 36979 |   | 36983 |
	| 15 GB   |       8 | 46087 |   | 49977 |
	| 15 GB   |      12 | 41901 |   | 48591 |
	| 15 GB   |      16 | 40147 |   | 50651 | +26.16%
	| 15 GB   |      24 | 37250 |   | 52365 | +40.58%
	| 15 GB   |      32 | 36470 |   | 50015 | +37.14%
Signed-off-by: default avatarMichael Wang <wangyun@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: default avatarPeter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/51D50057.9000809@linux.vnet.ibm.com
[ Improved the changelog. ]
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      62470419
    • Vladimir Davydov's avatar
      sched: Move h_load calculation to task_h_load() · 68520796
      Vladimir Davydov authored 
      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: default avatarVladimir Davydov <vdavydov@parallels.com>
Signed-off-by: default avatarPeter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1373896159-1278-1-git-send-email-vdavydov@parallels.comSigned-off-by: default avatarIngo Molnar <mingo@kernel.org>
      68520796
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7