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: Michael Wang <wangyun@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/51D50057.9000809@linux.vnet.ibm.com
[ Improved the changelog. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
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 a35b646
('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.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The __cpuinit type of throwaway sections might have made sense
some time ago when RAM was more constrained, but now the savings
do not offset the cost and complications. For example, the fix in
commit 5e427ec2d0 ("x86: Fix bit corruption at CPU resume time")
is a good example of the nasty type of bugs that can be created
with improper use of the various __init prefixes.
After a discussion on LKML[1] it was decided that cpuinit should go
the way of devinit and be phased out. Once all the users are gone,
we can then finally remove the macros themselves from linux/init.h.
This removes all the uses of the __cpuinit macros from C files in
the core kernel directories (kernel, init, lib, mm, and include)
that don't really have a specific maintainer.
[1] https://lkml.org/lkml/2013/5/20/589
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
David reported that the HRTICK sched feature was borken; which was enough
motivation for me to finally fix it ;-)
We should not allow hrtimer code to do softirq wakeups while holding scheduler
locks. The hrtimer code only needs this when we accidentally try to program an
expired time. We don't much care about those anyway since we have the regular
tick to fall back to.
Reported-by: David Ahern <dsahern@gmail.com>
Tested-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20130628091853.GE29209@dyad.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull timer core updates from Thomas Gleixner:
"The timer changes contain:
- posix timer code consolidation and fixes for odd corner cases
- sched_clock implementation moved from ARM to core code to avoid
duplication by other architectures
- alarm timer updates
- clocksource and clockevents unregistration facilities
- clocksource/events support for new hardware
- precise nanoseconds RTC readout (Xen feature)
- generic support for Xen suspend/resume oddities
- the usual lot of fixes and cleanups all over the place
The parts which touch other areas (ARM/XEN) have been coordinated with
the relevant maintainers. Though this results in an handful of
trivial to solve merge conflicts, which we preferred over nasty cross
tree merge dependencies.
The patches which have been committed in the last few days are bug
fixes plus the posix timer lot. The latter was in akpms queue and
next for quite some time; they just got forgotten and Frederic
collected them last minute."
* 'timers-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (59 commits)
hrtimer: Remove unused variable
hrtimers: Move SMP function call to thread context
clocksource: Reselect clocksource when watchdog validated high-res capability
posix-cpu-timers: don't account cpu timer after stopped thread runtime accounting
posix_timers: fix racy timer delta caching on task exit
posix-timers: correctly get dying task time sample in posix_cpu_timer_schedule()
selftests: add basic posix timers selftests
posix_cpu_timers: consolidate expired timers check
posix_cpu_timers: consolidate timer list cleanups
posix_cpu_timer: consolidate expiry time type
tick: Sanitize broadcast control logic
tick: Prevent uncontrolled switch to oneshot mode
tick: Make oneshot broadcast robust vs. CPU offlining
x86: xen: Sync the CMOS RTC as well as the Xen wallclock
x86: xen: Sync the wallclock when the system time is set
timekeeping: Indicate that clock was set in the pvclock gtod notifier
timekeeping: Pass flags instead of multiple bools to timekeeping_update()
xen: Remove clock_was_set() call in the resume path
hrtimers: Support resuming with two or more CPUs online (but stopped)
timer: Fix jiffies wrap behavior of round_jiffies_common()
...
When tsk->signal->cputimer->running is 1, signal->cputimer (i.e. per process
timer account) and tsk->sum_sched_runtime (i.e. per thread timer account)
increase at the same pace because update_curr() increases both accounting.
However, there is one exception. When thread exiting, __exit_signal() turns
over task's sum_shced_runtime to sig->sum_sched_runtime, but it doesn't stop
signal->cputimer accounting.
This inconsistency makes POSIX timer wake up too early. This patch fixes it.
Original-patch-by: Olivier Langlois <olivier@trillion01.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Olivier Langlois <olivier@trillion01.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Merge in a recent upstream commit:
c2853c8df5 include/linux/math64.h: add div64_ul()
because:
72a4cf20cb sched: Change cfs_rq load avg to unsigned long
relies on it.
[ We don't rebase sched/core for this, because the handful of
followup commits after the broken commit are not behavioral
changes so are unlikely to be needed during bisection. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
At present we print per-entity load-tracking statistics for
cfs_rq of cgroups/runqueues. Given that per task statistics
is maintained, it can be used to know the contribution made
by the task to its parenting cfs_rq level.
This patch adds per-task load-tracking statistics to /proc/<PID>/sched.
Signed-off-by: Kamalesh Babulal <kamalesh@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20130625080336.GA20175@linux.vnet.ibm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
They are the base values in load balance, update them with rq runnable
load average, then the load balance will consider runnable load avg
naturally.
We also try to include the blocked_load_avg as cpu load in balancing,
but that cause kbuild performance drop 6% on every Intel machine, and
aim7/oltp drop on some of 4 CPU sockets machines.
Or only add blocked_load_avg into get_rq_runable_load, hackbench still
drop a little on NHM EX.
Signed-off-by: Alex Shi <alex.shi@intel.com>
Reviewed-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1371694737-29336-7-git-send-email-alex.shi@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The woken migrated task will __synchronize_entity_decay(se); in
migrate_task_rq_fair, then it needs to set
`se->avg.last_runnable_update -= (-se->avg.decay_count) << 20' before
update_entity_load_avg, in order to avoid sleep time is updated twice
for se.avg.load_avg_contrib in both __syncchronize and
update_entity_load_avg.
However if the sleeping task is woken up from the same cpu, it miss
the last_runnable_update before update_entity_load_avg(se, 0, 1), then
the sleep time was used twice in both functions. So we need to remove
the double sleep time accounting.
Paul also contributed some code comments in this commit.
Signed-off-by: Alex Shi <alex.shi@intel.com>
Reviewed-by: Paul Turner <pjt@google.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1371694737-29336-5-git-send-email-alex.shi@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We need to initialize the se.avg.{decay_count, load_avg_contrib} for a
new forked task. Otherwise random values of above variables cause a
mess when a new task is enqueued:
enqueue_task_fair
enqueue_entity
enqueue_entity_load_avg
and make fork balancing imbalance due to incorrect load_avg_contrib.
Further more, Morten Rasmussen notice some tasks were not launched at
once after created. So Paul and Peter suggest giving a start value for
new task runnable avg time same as sched_slice().
PeterZ said:
> So the 'problem' is that our running avg is a 'floating' average; ie. it
> decays with time. Now we have to guess about the future of our newly
> spawned task -- something that is nigh impossible seeing these CPU
> vendors keep refusing to implement the crystal ball instruction.
>
> So there's two asymptotic cases we want to deal well with; 1) the case
> where the newly spawned program will be 'nearly' idle for its lifetime;
> and 2) the case where its cpu-bound.
>
> Since we have to guess, we'll go for worst case and assume its
> cpu-bound; now we don't want to make the avg so heavy adjusting to the
> near-idle case takes forever. We want to be able to quickly adjust and
> lower our running avg.
>
> Now we also don't want to make our avg too light, such that it gets
> decremented just for the new task not having had a chance to run yet --
> even if when it would run, it would be more cpu-bound than not.
>
> So what we do is we make the initial avg of the same duration as that we
> guess it takes to run each task on the system at least once -- aka
> sched_slice().
>
> Of course we can defeat this with wakeup/fork bombs, but in the 'normal'
> case it should be good enough.
Paul also contributed most of the code comments in this commit.
Signed-off-by: Alex Shi <alex.shi@intel.com>
Reviewed-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Reviewed-by: Paul Turner <pjt@google.com>
[peterz; added explanation of sched_slice() usage]
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1371694737-29336-4-git-send-email-alex.shi@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull scheduler fixes from Ingo Molnar:
"Two smaller fixes - plus a context tracking tracing fix that is a bit
bigger"
* 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
tracing/context-tracking: Add preempt_schedule_context() for tracing
sched: Fix clear NOHZ_BALANCE_KICK
sched/x86: Construct all sibling maps if smt
Peter Zijlstra
Michael Wang
Vladimir Davydov
Paul Gortmaker
Linus Torvalds
KOSAKI Motohiro
Ingo Molnar
Alex Shi
Kamalesh Babulal