When SD_PREFER_SIBLING is set on a sched domain, drop group_capacity to 1
only if the local group has extra capacity. The extra check prevents the case
where you always pull from the heaviest group when it is already under-utilized
(possible with a large weight task outweighs the tasks on the system).
For example, consider a 16-cpu quad-core quad-socket machine with MC and NUMA
scheduling domains. Let's say we spawn 15 nice0 tasks and one nice-15 task,
and each task is running on one core. In this case, we observe the following
events when balancing at the NUMA domain:
- find_busiest_group() will always pick the sched group containing the niced
task to be the busiest group.
- find_busiest_queue() will then always pick one of the cpus running the
nice0 task (never picks the cpu with the nice -15 task since
weighted_cpuload > imbalance).
- The load balancer fails to migrate the task since it is the running task
and increments sd->nr_balance_failed.
- It repeats the above steps a few more times until sd->nr_balance_failed > 5,
at which point it kicks off the active load balancer, wakes up the migration
thread and kicks the nice 0 task off the cpu.
The load balancer doesn't stop until we kick out all nice 0 tasks from
the sched group, leaving you with 3 idle cpus and one cpu running the
nice -15 task.
When balancing at the NUMA domain, we drop sgs.group_capacity to 1 if the child
domain (in this case MC) has SD_PREFER_SIBLING set. Subsequent load checks are
not relevant because the niced task has a very large weight.
In this patch, we add an extra condition to the "if(prefer_sibling)" check in
update_sd_lb_stats(). We drop the capacity of a group only if the local group
has extra capacity, ie. nr_running < group_capacity. This patch preserves the
original intent of the prefer_siblings check (to spread tasks across the system
in low utilization scenarios) and fixes the case above.
It helps in the following ways:
- In low utilization cases (where nr_tasks << nr_cpus), we still drop
group_capacity down to 1 if we prefer siblings.
- On very busy systems (where nr_tasks >> nr_cpus), sgs.nr_running will most
likely be > sgs.group_capacity.
- When balancing large weight tasks, if the local group does not have extra
capacity, we do not pick the group with the niced task as the busiest group.
This prevents failed balances, active migration and the under-utilization
described above.
Signed-off-by: Nikhil Rao <ncrao@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1287173550-30365-5-git-send-email-ncrao@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch forces a load balance on a newly idle cpu when the local group has
extra capacity and the busiest group does not have any. It improves system
utilization when balancing tasks with a large weight differential.
Under certain situations, such as a niced down task (i.e. nice = -15) in the
presence of nr_cpus NICE0 tasks, the niced task lands on a sched group and
kicks away other tasks because of its large weight. This leads to sub-optimal
utilization of the machine. Even though the sched group has capacity, it does
not pull tasks because sds.this_load >> sds.max_load, and f_b_g() returns NULL.
With this patch, if the local group has extra capacity, we shortcut the checks
in f_b_g() and try to pull a task over. A sched group has extra capacity if the
group capacity is greater than the number of running tasks in that group.
Thanks to Mike Galbraith for discussions leading to this patch and for the
insight to reuse SD_NEWIDLE_BALANCE.
Signed-off-by: Nikhil Rao <ncrao@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1287173550-30365-4-git-send-email-ncrao@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When cycling through sched groups to determine the busiest group, set
group_imb only if the busiest cpu has more than 1 runnable task. This patch
fixes the case where two cpus in a group have one runnable task each, but there
is a large weight differential between these two tasks. The load balancer is
unable to migrate any task from this group, and hence do not consider this
group to be imbalanced.
Signed-off-by: Nikhil Rao <ncrao@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1286996978-7007-3-git-send-email-ncrao@google.com>
[ small code readability edits ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
There's a situation where the nohz balancer will try to wake itself:
cpu-x is idle which is also ilb_cpu
got a scheduler tick during idle
and the nohz_kick_needed() in trigger_load_balance() checks for
rq_x->nr_running which might not be zero (because of someone waking a
task on this rq etc) and this leads to the situation of the cpu-x
sending a kick to itself.
And this can cause a lockup.
Avoid this by not marking ourself eligible for kicking.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1284400941.2684.19.camel@sbsiddha-MOBL3.sc.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Mathieu reported bad latencies with make -j10 kind of kbuild
workloads - which is mostly caused by us scheduling with a
too coarse granularity.
Reduce the minimum granularity some more, to make sure we
can meet the latency target.
I got the following results (make -j10 kbuild load, average of 3
runs):
vanilla:
maximum latency: 38278.9 µs
average latency: 7730.1 µs
patched:
maximum latency: 22702.1 µs
average latency: 6684.8 µs
Mathieu also measured it:
|
| * wakeup-latency.c (SIGEV_THREAD) with make -j10
|
| - Mainline 2.6.35.2 kernel
|
| maximum latency: 45762.1 µs
| average latency: 7348.6 µs
|
| - With only Peter's smaller min_gran (shown below):
|
| maximum latency: 29100.6 µs
| average latency: 6684.1 µs
|
Reported-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Suggested-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <AANLkTi=8m4g01wZPacySoF7U0PevTNVgJoZZrHiUD-pN@mail.gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
x86, tsc: Fix a preemption leak in restore_sched_clock_state()
sched: Move sched_avg_update() to update_cpu_load()
Currently sched_avg_update() (which updates rt_avg stats in the rq)
is getting called from scale_rt_power() (in the load balance context)
which doesn't take rq->lock.
Fix it by moving the sched_avg_update() to more appropriate
update_cpu_load() where the CFS load gets updated as well.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1282596171.2694.3.camel@sbsiddha-MOBL3>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
sched_fork() -- we do task placement in ->task_fork_fair() ensure we
update_rq_clock() so we work with current time. We leave the vruntime
in relative state, so the time delay until wake_up_new_task() doesn't
matter.
wake_up_new_task() -- Since task_fork_fair() left p->vruntime in
relative state we can safely migrate, the activate_task() on the
remote rq will call update_rq_clock() and causes the clock to be
synced (enough).
Tested-by: Jack Daniel <wanders.thirst@gmail.com>
Tested-by: Philby John <pjohn@mvista.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1281002322.1923.1708.camel@laptop>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently we update cpu_power() too often, update_group_power() only
updates the local group's cpu_power but it gets called for all groups.
Furthermore, CPU_NEWLY_IDLE invocations will result in all cpus
calling it, even though a slow update of cpu_power is sufficient.
Therefore move the update under 'idle != CPU_NEWLY_IDLE &&
local_group' to reduce superfluous invocations.
Reported-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Suresh Siddha <suresh.b.siddha@intel.com>
LKML-Reference: <1278612989.1900.176.camel@laptop>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The task_group() function returns a pointer that must be protected
by either RCU, the ->alloc_lock, or the cgroup lock (see the
rcu_dereference_check() in task_subsys_state(), which is invoked by
task_group()). The wake_affine() function currently does none of these,
which means that a concurrent update would be within its rights to free
the structure returned by task_group(). Because wake_affine() uses this
structure only to compute load-balancing heuristics, there is no reason
to acquire either of the two locks.
Therefore, this commit introduces an RCU read-side critical section that
starts before the first call to task_group() and ends after the last use
of the "tg" pointer returned from task_group(). Thanks to Li Zefan for
pointing out the need to extend the RCU read-side critical section from
that proposed by the original patch.
Signed-off-by: Daniel J Blueman <daniel.blueman@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Docbook fails in sched_fair.c due to comments added in the asymmetric
packing patch series.
This fixes these errors. No code changes.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <24737.1276135581@neuling.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The CPU power test is the wrong way around in fix_small_capacity.
This was due to a small changes made in the posted patch on lkml to what
was was taken upstream.
This patch fixes asymmetric packing for POWER7.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <12629.1276124617@neuling.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Check to see if the group is packed in a sched doman.
This is primarily intended to used at the sibling level. Some cores
like POWER7 prefer to use lower numbered SMT threads. In the case of
POWER7, it can move to lower SMT modes only when higher threads are
idle. When in lower SMT modes, the threads will perform better since
they share less core resources. Hence when we have idle threads, we
want them to be the higher ones.
This adds a hook into f_b_g() called check_asym_packing() to check the
packing. This packing function is run on idle threads. It checks to
see if the busiest CPU in this domain (core in the P7 case) has a
higher CPU number than what where the packing function is being run
on. If it is, calculate the imbalance and return the higher busier
thread as the busiest group to f_b_g(). Here we are assuming a lower
CPU number will be equivalent to a lower SMT thread number.
It also creates a new SD_ASYM_PACKING flag to enable this feature at
any scheduler domain level.
It also creates an arch hook to enable this feature at the sibling
level. The default function doesn't enable this feature.
Based heavily on patch from Peter Zijlstra.
Fixes from Srivatsa Vaddagiri.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <20100608045702.2936CCC897@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Handle cpu capacity being reported as 0 on cores with more number of
hardware threads. For example on a Power7 core with 4 hardware
threads, core power is 1177 and thus power of each hardware thread is
1177/4 = 294. This low power can lead to capacity for each hardware
thread being calculated as 0, which leads to tasks bouncing within the
core madly!
Fix this by reporting capacity for hardware threads as 1, provided
their power is not scaled down significantly because of frequency
scaling or real-time tasks usage of cpu.
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arjan van de Ven <arjan@linux.intel.com>
LKML-Reference: <20100608045702.21D03CC895@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In the new push model, all idle CPUs indeed go into nohz mode. There is
still the concept of idle load balancer (performing the load balancing
on behalf of all the idle cpu's in the system). Busy CPU kicks the nohz
balancer when any of the nohz CPUs need idle load balancing.
The kickee CPU does the idle load balancing on behalf of all idle CPUs
instead of the normal idle balance.
This addresses the below two problems with the current nohz ilb logic:
* the idle load balancer continued to have periodic ticks during idle and
wokeup frequently, even though it did not have any rebalancing to do on
behalf of any of the idle CPUs.
* On x86 and CPUs that have APIC timer stoppage on idle CPUs, this
periodic wakeup can result in a periodic additional interrupt on a CPU
doing the timer broadcast.
Also currently we are migrating the unpinned timers from an idle to the cpu
doing idle load balancing (when all the cpus in the system are idle,
there is no idle load balancing cpu and timers get added to the same idle cpu
where the request was made. So the existing optimization works only on semi idle
system).
And In semi idle system, we no longer have periodic ticks on the idle load
balancer CPU. Using that cpu will add more delays to the timers than intended
(as that cpu's timer base may not be uptodate wrt jiffies etc). This was
causing mysterious slowdowns during boot etc.
For now, in the semi idle case, use the nearest busy cpu for migrating timers
from an idle cpu. This is good for power-savings anyway.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <1274486981.2840.46.camel@sbs-t61.sc.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
tickless idle has a negative side effect on update_cpu_load(), which
in turn can affect load balancing behavior.
update_cpu_load() is supposed to be called every tick, to keep track
of various load indicies. With tickless idle, there are no scheduler
ticks called on the idle CPUs. Idle CPUs may still do load balancing
(with idle_load_balance CPU) using the stale cpu_load. It will also
cause problems when all CPUs go idle for a while and become active
again. In this case loads would not degrade as expected.
This is how rq->nr_load_updates change looks like under different
conditions:
<cpu_num> <nr_load_updates change>
All CPUS idle for 10 seconds (HZ=1000)
0 1621
10 496
11 139
12 875
13 1672
14 12
15 21
1 1472
2 2426
3 1161
4 2108
5 1525
6 701
7 249
8 766
9 1967
One CPU busy rest idle for 10 seconds
0 10003
10 601
11 95
12 966
13 1597
14 114
15 98
1 3457
2 93
3 6679
4 1425
5 1479
6 595
7 193
8 633
9 1687
All CPUs busy for 10 seconds
0 10026
10 10026
11 10026
12 10026
13 10025
14 10025
15 10025
1 10026
2 10026
3 10026
4 10026
5 10026
6 10026
7 10026
8 10026
9 10026
That is update_cpu_load works properly only when all CPUs are busy.
If all are idle, all the CPUs get way lower updates. And when few
CPUs are busy and rest are idle, only busy and ilb CPU does proper
updates and rest of the idle CPUs will do lower updates.
The patch keeps track of when a last update was done and fixes up
the load avg based on current time.
On one of my test system SPECjbb with warehouse 1..numcpus, patch
improves throughput numbers by ~1% (average of 6 runs). On another
test system (with different domain hierarchy) there is no noticable
change in perf.
Signed-off-by: Venkatesh Pallipadi <venki@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <AANLkTilLtDWQsAUrIxJ6s04WTgmw9GuOODc5AOrYsaR5@mail.gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Mike reports that since e9e9250b (sched: Scale down cpu_power due to RT
tasks), wake_affine() goes funny on RT tasks due to them still having a
!0 weight and wake_affine() still subtracts that from the rq weight.
Since nobody should be using se->weight for RT tasks, set the value to
zero. Also, since we now use ->cpu_power to normalize rq weights to
account for RT cpu usage, add that factor into the imbalance computation.
Reported-by: Mike Galbraith <efault@gmx.de>
Tested-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1275316109.27810.22969.camel@twins>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently migration_thread is serving three purposes - migration
pusher, context to execute active_load_balance() and forced context
switcher for expedited RCU synchronize_sched. All three roles are
hardcoded into migration_thread() and determining which job is
scheduled is slightly messy.
This patch kills migration_thread and replaces all three uses with
cpu_stop. The three different roles of migration_thread() are
splitted into three separate cpu_stop callbacks -
migration_cpu_stop(), active_load_balance_cpu_stop() and
synchronize_sched_expedited_cpu_stop() - and each use case now simply
asks cpu_stop to execute the callback as necessary.
synchronize_sched_expedited() was implemented with private
preallocated resources and custom multi-cpu queueing and waiting
logic, both of which are provided by cpu_stop.
synchronize_sched_expedited_count is made atomic and all other shared
resources along with the mutex are dropped.
synchronize_sched_expedited() also implemented a check to detect cases
where not all the callback got executed on their assigned cpus and
fall back to synchronize_sched(). If called with cpu hotplug blocked,
cpu_stop already guarantees that and the condition cannot happen;
otherwise, stop_machine() would break. However, this patch preserves
the paranoid check using a cpumask to record on which cpus the stopper
ran so that it can serve as a bisection point if something actually
goes wrong theree.
Because the internal execution state is no longer visible,
rcu_expedited_torture_stats() is removed.
This patch also renames cpu_stop threads to from "stopper/%d" to
"migration/%d". The names of these threads ultimately don't matter
and there's no reason to make unnecessary userland visible changes.
With this patch applied, stop_machine() and sched now share the same
resources. stop_machine() is faster without wasting any resources and
sched migration users are much cleaner.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Dipankar Sarma <dipankar@in.ibm.com>
Cc: Josh Triplett <josh@freedesktop.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Dimitri Sivanich <sivanich@sgi.com>
Nikhil Rao
Takuya Yoshikawa
Ingo Molnar
Venkatesh Pallipadi
Suresh Siddha
Linus Torvalds
Andi Kleen
Peter Zijlstra
Peter Zijlstra
Michael Neuling
Daniel J Blueman
Srivatsa Vaddagiri
Tejun Heo