Excessive migration of pages can hurt the performance of workloads
that span multiple NUMA nodes. However, it turns out that the
p->numa_migrate_deferred knob is a really big hammer, which does
reduce migration rates, but does not actually help performance.
Now that the second stage of the automatic numa balancing code
has stabilized, it is time to replace the simplistic migration
deferral code with something smarter.
Signed-off-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Chegu Vinod <chegu_vinod@hp.com>
Link: http://lkml.kernel.org/r/1390860228-21539-2-git-send-email-riel@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull scheduler fixes from Ingo Molnar:
"A couple of regression fixes mostly hitting virtualized setups, but
also some bare metal systems"
* 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/x86/tsc: Initialize multiplier to 0
sched/clock: Fixup early initialization
sched/preempt/x86: Fix voluntary preempt for x86
Revert "sched: Fix sleep time double accounting in enqueue entity"
This reverts commit 282cf499f0.
With the current implementation, the load average statistics of a sched entity
change according to other activity on the CPU even if this activity is done
between the running window of the sched entity and have no influence on the
running duration of the task.
When a task wakes up on the same CPU, we currently update last_runnable_update
with the return of __synchronize_entity_decay without updating the
runnable_avg_sum and runnable_avg_period accordingly. In fact, we have to sync
the load_contrib of the se with the rq's blocked_load_contrib before removing
it from the latter (with __synchronize_entity_decay) but we must keep
last_runnable_update unchanged for updating runnable_avg_sum/period during the
next update_entity_load_avg.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Ben Segall <bsegall@google.com>
Cc: pjt@google.com
Cc: alex.shi@linaro.org
Link: http://lkml.kernel.org/r/1390376734-6800-1-git-send-email-vincent.guittot@linaro.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch adds three tracepoints
o trace_sched_move_numa when a task is moved to a node
o trace_sched_swap_numa when a task is swapped with another task
o trace_sched_stick_numa when a numa-related migration fails
The tracepoints allow the NUMA scheduler activity to be monitored and the
following high-level metrics can be calculated
o NUMA migrated stuck nr trace_sched_stick_numa
o NUMA migrated idle nr trace_sched_move_numa
o NUMA migrated swapped nr trace_sched_swap_numa
o NUMA local swapped trace_sched_swap_numa src_nid == dst_nid (should never happen)
o NUMA remote swapped trace_sched_swap_numa src_nid != dst_nid (should == NUMA migrated swapped)
o NUMA group swapped trace_sched_swap_numa src_ngid == dst_ngid
Maybe a small number of these are acceptable
but a high number would be a major surprise.
It would be even worse if bounces are frequent.
o NUMA avg task migs. Average number of migrations for tasks
o NUMA stddev task mig Self-explanatory
o NUMA max task migs. Maximum number of migrations for a single task
In general the intent of the tracepoints is to help diagnose problems
where automatic NUMA balancing appears to be doing an excessive amount
of useless work.
[akpm@linux-foundation.org: remove semicolon-after-if, repair coding-style]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Alex Thorlton <athorlton@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Thomas Hellstrom bisected a regression where erratic 3D performance is
experienced on virtual machines as measured by glxgears. It identified
commit 58d081b5 ("sched/numa: Avoid overloading CPUs on a preferred NUMA
node") as the problem which had modified the behaviour of effective_load.
Effective load calculates the difference to the system-wide load if a
scheduling entity was moved to another CPU. The task group is not heavier
as a result of the move but overall system load can increase/decrease as a
result of the change. Commit 58d081b5 ("sched/numa: Avoid overloading CPUs
on a preferred NUMA node") changed effective_load to make it suitable for
calculating if a particular NUMA node was compute overloaded. To reduce
the cost of the function, it assumed that a current sched entity weight
of 0 was uninteresting but that is not the case.
wake_affine() uses a weight of 0 for sync wakeups on the grounds that it
is assuming the waking task will sleep and not contribute to load in the
near future. In this case, we still want to calculate the effective load
of the sched entity hierarchy. As effective_load is no longer used by
task_numa_compare since commit fb13c7ee (sched/numa: Use a system-wide
search to find swap/migration candidates), this patch simply restores the
historical behaviour.
Reported-and-tested-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
[ Wrote changelog]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20140106113912.GC6178@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This
results in him occasionally seeing time going backwards - which
crashes the scheduler ...
Most of our time accounting can actually handle that except the most
common one; the tick time update of sched_fair.
There is a further problem with that code; previously we assumed that
because we get a tick every TICK_NSEC our time delta could never
exceed 32bits and math was simpler.
However, ever since Frederic managed to get NO_HZ_FULL merged; this is
no longer the case since now a task can run for a long time indeed
without getting a tick. It only takes about ~4.2 seconds to overflow
our u32 in nanoseconds.
This means we not only need to better deal with time going backwards;
but also means we need to be able to deal with large deltas.
This patch reworks the entire code and uses mul_u64_u32_shr() as
proposed by Andy a long while ago.
We express our virtual time scale factor in a u32 multiplier and shift
right and the 32bit mul_u64_u32_shr() implementation reduces to a
single 32x32->64 multiply if the time delta is still short (common
case).
For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128.
Reported-and-Tested-by: Christian Engelmayer <cengelma@gmx.at>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: fweisbec@gmail.com
Cc: Paul Turner <pjt@google.com>
Cc: Stanislaw Gruszka <sgruszka@redhat.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Because we're completely unserialized against hotplug its well
possible to try and generate numa stats for an offlined node.
Bail out early (and avoid a /0) in this case. The resulting stats are
all 0 which should result in an undesirable balance target -- not to
mention that actually trying to migrate to an offline CPU will fail.
Reported-by: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Link: http://lkml.kernel.org/n/tip-orja0qylcvyhxfsuebcyL5sI@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The cpusets code can split up the scheduler's domain tree into
smaller domains. Some of those smaller domains may not cross
NUMA nodes at all, leading to a NULL pointer dereference on the
per-cpu sd_numa pointer.
Tasks cannot be migrated out of their domain, so the patch
also sets p->numa_preferred_nid to whereever they are, to
prevent the migration from being retried over and over again.
Reported-by: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Link: http://lkml.kernel.org/n/tip-oosqomw0Jput0Jkvoowhrqtu@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
nr_busy_cpus parameter is used by nohz_kick_needed() to find out the
number of busy cpus in a sched domain which has SD_SHARE_PKG_RESOURCES
flag set. Therefore instead of updating nr_busy_cpus at every level
of sched domain, since it is irrelevant, we can update this parameter
only at the parent domain of the sd which has this flag set. Introduce
a per-cpu parameter sd_busy which represents this parent domain.
In nohz_kick_needed() we directly query the nr_busy_cpus parameter
associated with the groups of sd_busy.
By associating sd_busy with the highest domain which has
SD_SHARE_PKG_RESOURCES flag set, we cover all lower level domains
which could have this flag set and trigger nohz_idle_balancing if any
of the levels have more than one busy cpu.
sd_busy is irrelevant for asymmetric load balancing. However sd_asym
has been introduced to represent the highest sched domain which has
SD_ASYM_PACKING flag set so that it can be queried directly when
required.
While we are at it, we might as well change the nohz_idle parameter to
be updated at the sd_busy domain level alone and not the base domain
level of a CPU. This will unify the concept of busy cpus at just one
level of sched domain where it is currently used.
Signed-off-by: Preeti U Murthy<preeti@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: svaidy@linux.vnet.ibm.com
Cc: vincent.guittot@linaro.org
Cc: bitbucket@online.de
Cc: benh@kernel.crashing.org
Cc: anton@samba.org
Cc: Morten.Rasmussen@arm.com
Cc: pjt@google.com
Cc: peterz@infradead.org
Cc: mikey@neuling.org
Link: http://lkml.kernel.org/r/20131030031252.23426.4417.stgit@preeti.in.ibm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
throttle_cfs_rq() doesn't check to make sure that period_timer is running,
and while update_curr/assign_cfs_runtime does, a concurrently running
period_timer on another cpu could cancel itself between this cpu's
update_curr and throttle_cfs_rq(). If there are no other cfs_rqs running
in the tg to restart the timer, this causes the cfs_rq to be stranded
forever.
Fix this by calling __start_cfs_bandwidth() in throttle if the timer is
inactive.
(Also add some sched_debug lines for cfs_bandwidth.)
Tested: make a run/sleep task in a cgroup, loop switching the cgroup
between 1ms/100ms quota and unlimited, checking for timer_active=0 and
throttled=1 as a failure. With the throttle_cfs_rq() change commented out
this fails, with the full patch it passes.
Signed-off-by: Ben Segall <bsegall@google.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: pjt@google.com
Link: http://lkml.kernel.org/r/20131016181632.22647.84174.stgit@sword-of-the-dawn.mtv.corp.google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Currently, group entity load-weights are initialized to zero. This
admits some races with respect to the first time they are re-weighted in
earlty use. ( Let g[x] denote the se for "g" on cpu "x". )
Suppose that we have root->a and that a enters a throttled state,
immediately followed by a[0]->t1 (the only task running on cpu[0])
blocking:
put_prev_task(group_cfs_rq(a[0]), t1)
put_prev_entity(..., t1)
check_cfs_rq_runtime(group_cfs_rq(a[0]))
throttle_cfs_rq(group_cfs_rq(a[0]))
Then, before unthrottling occurs, let a[0]->b[0]->t2 wake for the first
time:
enqueue_task_fair(rq[0], t2)
enqueue_entity(group_cfs_rq(b[0]), t2)
enqueue_entity_load_avg(group_cfs_rq(b[0]), t2)
account_entity_enqueue(group_cfs_ra(b[0]), t2)
update_cfs_shares(group_cfs_rq(b[0]))
< skipped because b is part of a throttled hierarchy >
enqueue_entity(group_cfs_rq(a[0]), b[0])
...
We now have b[0] enqueued, yet group_cfs_rq(a[0])->load.weight == 0
which violates invariants in several code-paths. Eliminate the
possibility of this by initializing group entity weight.
Signed-off-by: Paul Turner <pjt@google.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20131016181627.22647.47543.stgit@sword-of-the-dawn.mtv.corp.google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
__start_cfs_bandwidth calls hrtimer_cancel while holding rq->lock,
waiting for the hrtimer to finish. However, if sched_cfs_period_timer
runs for another loop iteration, the hrtimer can attempt to take
rq->lock, resulting in deadlock.
Fix this by ensuring that cfs_b->timer_active is cleared only if the
_latest_ call to do_sched_cfs_period_timer is returning as idle. Then
__start_cfs_bandwidth can just call hrtimer_try_to_cancel and wait for
that to succeed or timer_active == 1.
Signed-off-by: Ben Segall <bsegall@google.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: pjt@google.com
Link: http://lkml.kernel.org/r/20131016181622.22647.16643.stgit@sword-of-the-dawn.mtv.corp.google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When we transition cfs_bandwidth_used to false, any currently
throttled groups will incorrectly return false from cfs_rq_throttled.
While tg_set_cfs_bandwidth will unthrottle them eventually, currently
running code (including at least dequeue_task_fair and
distribute_cfs_runtime) will cause errors.
Fix this by turning off cfs_bandwidth_used only after unthrottling all
cfs_rqs.
Tested: toggle bandwidth back and forth on a loaded cgroup. Caused
crashes in minutes without the patch, hasn't crashed with it.
Signed-off-by: Ben Segall <bsegall@google.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: pjt@google.com
Link: http://lkml.kernel.org/r/20131016181611.22647.80365.stgit@sword-of-the-dawn.mtv.corp.google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
There is a subtle race in migrate_swap, when task P, on CPU A, decides to swap
places with task T, on CPU B.
Task P:
- call migrate_swap
Task T:
- go to sleep, removing itself from the runqueue
Task P:
- double lock the runqueues on CPU A & B
Task T:
- get woken up, place itself on the runqueue of CPU C
Task P:
- see that task T is on a runqueue, and pretend to remove it
from the runqueue on CPU B
Now CPUs B & C both have corrupted scheduler data structures.
This patch fixes it, by holding the pi_lock for both of the tasks
involved in the migrate swap. This prevents task T from waking up,
and placing itself onto another runqueue, until after migrate_swap
has released all locks.
This means that, when migrate_swap checks, task T will be either
on the runqueue where it was originally seen, or not on any
runqueue at all. Migrate_swap deals correctly with of those cases.
Tested-by: Joe Mario <jmario@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: hannes@cmpxchg.org
Cc: aarcange@redhat.com
Cc: srikar@linux.vnet.ibm.com
Cc: tglx@linutronix.de
Cc: hpa@zytor.com
Link: http://lkml.kernel.org/r/20131010181722.GO13848@laptop.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reflow the function a bit because GCC gets confused:
kernel/sched/fair.c: In function ‘task_numa_fault’:
kernel/sched/fair.c:1448:3: warning: ‘my_grp’ may be used uninitialized in this function [-Wmaybe-uninitialized]
kernel/sched/fair.c:1463:27: note: ‘my_grp’ was declared here
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/n/tip-6ebt6x7u64pbbonq1khqu2z9@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Short spikes of CPU load can lead to a task being migrated
away from its preferred node for temporary reasons.
It is important that the task is migrated back to where it
belongs, in order to avoid migrating too much memory to its
new location, and generally disturbing a task's NUMA location.
This patch fixes NUMA placement for 4 specjbb instances on
a 4 node system. Without this patch, things take longer to
converge, and processes are not always completely on their
own node.
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-64-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Shared faults can lead to lots of unnecessary page migrations,
slowing down the system, and causing private faults to hit the
per-pgdat migration ratelimit.
This patch adds sysctl numa_balancing_migrate_deferred, which specifies
how many shared page migrations to skip unconditionally, after each page
migration that is skipped because it is a shared fault.
This reduces the number of page migrations back and forth in
shared fault situations. It also gives a strong preference to
the tasks that are already running where most of the memory is,
and to moving the other tasks to near the memory.
Testing this with a much higher scan rate than the default
still seems to result in fewer page migrations than before.
Memory seems to be somewhat better consolidated than previously,
with multi-instance specjbb runs on a 4 node system.
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-62-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Adjust numa_scan_period in task_numa_placement, depending on how much
useful work the numa code can do. The more local faults there are in a
given scan window the longer the period (and hence the slower the scan rate)
during the next window. If there are excessive shared faults then the scan
period will decrease with the amount of scaling depending on whether the
ratio of shared/private faults. If the preferred node changes then the
scan rate is reset to recheck if the task is properly placed.
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-59-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch classifies scheduler domains and runqueues into types depending
the number of tasks that are about their NUMA placement and the number
that are currently running on their preferred node. The types are
regular: There are tasks running that do not care about their NUMA
placement.
remote: There are tasks running that care about their placement but are
currently running on a node remote to their ideal placement
all: No distinction
To implement this the patch tracks the number of tasks that are optimally
NUMA placed (rq->nr_preferred_running) and the number of tasks running
that care about their placement (nr_numa_running). The load balancer
uses this information to avoid migrating idea placed NUMA tasks as long
as better options for load balancing exists. For example, it will not
consider balancing between a group whose tasks are all perfectly placed
and a group with remote tasks.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/1381141781-10992-56-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
