find_idlest_group() returns NULL when the local group is idlest. The
caller then continues the find_idlest_group() search at a lower level
of the current CPU's sched_domain hierarchy. find_idlest_group_cpu() is
not consulted and, crucially, @new_cpu is not updated. This means the
search is pointless and we return @prev_cpu from select_task_rq_fair().
This is fixed by initialising @new_cpu to @cpu instead of @prev_cpu.
Signed-off-by: Brendan Jackman <brendan.jackman@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20171005114516.18617-6-brendan.jackman@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When 'p' is not allowed on any of the CPUs in the sched_domain, we
currently return NULL from find_idlest_group(), and pointlessly
continue the search on lower sched_domain levels (where 'p' is also not
allowed) before returning prev_cpu regardless (as we have not updated
new_cpu).
Add an explicit check for this case, and add a comment to
find_idlest_group(). Now when find_idlest_group() returns NULL, it always
means that the local group is allowed and idlest.
Signed-off-by: Brendan Jackman <brendan.jackman@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20171005114516.18617-5-brendan.jackman@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When the local group is not allowed we do not modify this_*_load from
their initial value of 0. That means that the load checks at the end
of find_idlest_group cause us to incorrectly return NULL. Fixing the
initial values to ULONG_MAX means we will instead return the idlest
remote group in that case.
Signed-off-by: Brendan Jackman <brendan.jackman@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20171005114516.18617-4-brendan.jackman@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Since commit:
83a0a96a5f ("sched/fair: Leverage the idle state info when choosing the "idlest" cpu")
find_idlest_group_cpu() (formerly find_idlest_cpu) no longer returns -1,
so we can simplify the checking of the return value in find_idlest_cpu().
Signed-off-by: Brendan Jackman <brendan.jackman@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20171005114516.18617-3-brendan.jackman@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In preparation for changes that would otherwise require adding a new
level of indentation to the while(sd) loop, create a new function
find_idlest_cpu() which contains this loop, and rename the existing
find_idlest_cpu() to find_idlest_group_cpu().
Code inside the while(sd) loop is unchanged. @new_cpu is added as a
variable in the new function, with the same initial value as the
@new_cpu in select_task_rq_fair().
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Brendan Jackman <brendan.jackman@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20171005114516.18617-2-brendan.jackman@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The "goto force_balance" here is intended to mitigate the fact that
avg_load calculations can result in bad placement decisions when
priority is asymmetrical.
The original commit that adds it:
fab476228b ("sched: Force balancing on newidle balance if local group has capacity")
explains:
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.
A similar but inverted issue also affects ARM big.LITTLE (asymmetrical CPU
capacity) systems - consider 8 always-running, same-priority tasks on a
system with 4 "big" and 4 "little" CPUs. Suppose that 5 of them end up on
the "big" CPUs (which will be represented by one sched_group in the DIE
sched_domain) and 3 on the "little" (the other sched_group in DIE), leaving
one CPU unused. Because the "big" group has a higher group_capacity its
avg_load may not present an imbalance that would cause migrating a
task to the idle "little".
The force_balance case here solves the problem but currently only for
CPU_NEWLY_IDLE balances, which in theory might never happen on the
unused CPU. Including CPU_IDLE in the force_balance case means
there's an upper bound on the time before we can attempt to solve the
underutilization: after DIE's sd->balance_interval has passed the
next nohz balance kick will help us out.
Signed-off-by: Brendan Jackman <brendan.jackman@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Paul Turner <pjt@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20170807163900.25180-1-brendan.jackman@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We use task_util() in find_idlest_group() via capacity_spare_wake().
This task_util() updated in wake_cap(). However wake_cap() is not the
only reason for ending up in find_idlest_group() - we could have been sent
there by wake_wide(). So explicitly sync the task util with prev_cpu
when we are about to head to find_idlest_group().
We could simply do this at the beginning of
select_task_rq_fair() (i.e. irrespective of whether we're heading to
select_idle_sibling() or find_idlest_group() & co), but I didn't want to
slow down the select_idle_sibling() path more than necessary.
Don't do this during fork balancing, we won't need the task_util and
we'd just clobber the last_update_time, which is supposed to be 0.
Signed-off-by: Brendan Jackman <brendan.jackman@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andres Oportus <andresoportus@google.com>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Link: http://lkml.kernel.org/r/20170808095519.10077-1-brendan.jackman@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
As a first step this patch makes cfs_tasks list as MRU one.
It means, that when a next task is picked to run on physical
CPU it is moved to the front of the list.
Therefore, the cfs_tasks list is more or less sorted (except
woken tasks) starting from recently given CPU time tasks toward
tasks with max wait time in a run-queue, i.e. MRU list.
Second, as part of the load balance operation, this approach
starts detach_tasks()/detach_one_task() from the tail of the
queue instead of the head, giving some advantages:
- tends to pick a task with highest wait time;
- tasks located in the tail are less likely cache-hot,
therefore the can_migrate_task() decision is higher.
hackbench illustrates slightly better performance. For example
doing 1000 samples and 40 groups on i5-3320M CPU, it shows below
figures:
default: 0.657 avg
patched: 0.646 avg
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Kirill Tkhai <tkhai@yandex.ru>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Mike Galbraith <umgwanakikbuti@gmail.com>
Cc: Nicolas Pitre <nicolas.pitre@linaro.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Paul Turner <pjt@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Link: http://lkml.kernel.org/r/20170913102430.8985-2-urezki@gmail.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
On AMD Family17h-based (EPYC) system, a logical NUMA node can contain
upto 8 cores (16 threads) with the following topology.
----------------------------
C0 | T0 T1 | || | T0 T1 | C4
--------| || |--------
C1 | T0 T1 | L3 || L3 | T0 T1 | C5
--------| || |--------
C2 | T0 T1 | #0 || #1 | T0 T1 | C6
--------| || |--------
C3 | T0 T1 | || | T0 T1 | C7
----------------------------
Here, there are 2 last-level (L3) caches per logical NUMA node.
A socket can contain upto 4 NUMA nodes, and a system can support
upto 2 sockets. With full system configuration, current scheduler
creates 4 sched domains:
domain0 SMT (span a core)
domain1 MC (span a last-level-cache)
domain2 NUMA (span a socket: 4 nodes)
domain3 NUMA (span a system: 8 nodes)
Note that there is no domain to represent cpus spaning a logical
NUMA node. With this hierarchy of sched domains, the scheduler does
not balance properly in the following cases:
Case1:
When running 8 tasks, a properly balanced system should
schedule a task per logical NUMA node. This is not the case for
the current scheduler.
Case2:
In some cases, threads are scheduled on the same cpu, while other
cpus are idle. This results in run-to-run inconsistency. For example:
taskset -c 0-7 sysbench --num-threads=8 --test=cpu \
--cpu-max-prime=100000 run
Total execution time ranges from 25.1s to 33.5s depending on threads
placement, where 25.1s is when all 8 threads are balanced properly
on 8 cpus.
Introducing NUMA identity node sched domain, which is based on how
SRAT/SLIT table define a logical NUMA node. This results in the following
hierarchy of sched domains on the same system described above.
domain0 SMT (span a core)
domain1 MC (span a last-level-cache)
domain2 NODE (span a logical NUMA node)
domain3 NUMA (span a socket: 4 nodes)
domain4 NUMA (span a system: 8 nodes)
This fixes the improper load balancing cases mentioned above.
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: bp@suse.de
Link: http://lkml.kernel.org/r/1504768805-46716-1-git-send-email-suravee.suthikulpanit@amd.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The normal x86_topology on NHM+ machines degenerates because the MC
and CPU domains are of the same size, therefore MC inherits
SD_PREFER_SIBLING from CPU (which then gets taken out). The result is
that we'll spread tasks across the first NUMA level in order to
maximize cache utilization.
However, for the x86_numa_in_package_topology we loose the CPU domain,
and we'll not have SD_PREFER_SIBLING set anywhere, giving a distinct
difference in behaviour.
Commit:
8e7fbcbc22 ("sched: Remove stale power aware scheduling remnants and dysfunctional knobs")
made a fail by not preserving the SD_PREFER_SIBLING for the !power_saving
case on both CPU and MC.
Then commit:
6956dc568f ("sched/numa: Add SD_PERFER_SIBLING to CPU domain")
adds it back to the CPU but not MC.
Restore that now, such that we get consistent spreading behaviour wrt
L3 and NUMA.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
__dl_sub() is more meaningful as a name, and is more consistent
with the naming of the dual function (__dl_add()).
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Luca Abeni <luca.abeni@santannapisa.it>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Cc: Juri Lelli <juri.lelli@arm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1504778971-13573-4-git-send-email-luca.abeni@santannapisa.it
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Fix a bug introduced in:
72f9f3fdc9 ("sched/deadline: Remove dl_new from struct sched_dl_entity")
After that commit, when switching to -deadline if the scheduling
deadline of a task is in the past then switched_to_dl() calls
setup_new_entity() to properly initialize the scheduling deadline
and runtime.
The problem is that the task is enqueued _before_ having its parameters
initialized by setup_new_entity(), and this can cause problems.
For example, a task with its out-of-date deadline in the past will
potentially be enqueued as the highest priority one; however, its
adjusted deadline may not be the earliest one.
This patch fixes the problem by initializing the task's parameters before
enqueuing it.
Signed-off-by: luca abeni <luca.abeni@santannapisa.it>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Cc: Juri Lelli <juri.lelli@arm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1504778971-13573-3-git-send-email-luca.abeni@santannapisa.it
Signed-off-by: Ingo Molnar <mingo@kernel.org>
quiet_vmstat() is an expensive function that only makes sense when we
go into NOHZ.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: aubrey.li@linux.intel.com
Cc: cl@linux.com
Cc: fweisbec@gmail.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
While load_balance() masks the source CPUs against active_mask, it had
a hole against the destination CPU. Ensure the destination CPU is also
part of the 'domain-mask & active-mask' set.
Reported-by: Levin, Alexander (Sasha Levin) <alexander.levin@verizon.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Fixes: 77d1dfda0e ("sched/topology, cpuset: Avoid spurious/wrong domain rebuilds")
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The trivial wake_affine_idle() implementation is very good for a
number of workloads, but it comes apart at the moment there are no
idle CPUs left, IOW. the overloaded case.
hackbench:
NO_WA_WEIGHT WA_WEIGHT
hackbench-20 : 7.362717561 seconds 6.450509391 seconds
(win)
netperf:
NO_WA_WEIGHT WA_WEIGHT
TCP_SENDFILE-1 : Avg: 54524.6 Avg: 52224.3
TCP_SENDFILE-10 : Avg: 48185.2 Avg: 46504.3
TCP_SENDFILE-20 : Avg: 29031.2 Avg: 28610.3
TCP_SENDFILE-40 : Avg: 9819.72 Avg: 9253.12
TCP_SENDFILE-80 : Avg: 5355.3 Avg: 4687.4
TCP_STREAM-1 : Avg: 41448.3 Avg: 42254
TCP_STREAM-10 : Avg: 24123.2 Avg: 25847.9
TCP_STREAM-20 : Avg: 15834.5 Avg: 18374.4
TCP_STREAM-40 : Avg: 5583.91 Avg: 5599.57
TCP_STREAM-80 : Avg: 2329.66 Avg: 2726.41
TCP_RR-1 : Avg: 80473.5 Avg: 82638.8
TCP_RR-10 : Avg: 72660.5 Avg: 73265.1
TCP_RR-20 : Avg: 52607.1 Avg: 52634.5
TCP_RR-40 : Avg: 57199.2 Avg: 56302.3
TCP_RR-80 : Avg: 25330.3 Avg: 26867.9
UDP_RR-1 : Avg: 108266 Avg: 107844
UDP_RR-10 : Avg: 95480 Avg: 95245.2
UDP_RR-20 : Avg: 68770.8 Avg: 68673.7
UDP_RR-40 : Avg: 76231 Avg: 75419.1
UDP_RR-80 : Avg: 34578.3 Avg: 35639.1
UDP_STREAM-1 : Avg: 64684.3 Avg: 66606
UDP_STREAM-10 : Avg: 52701.2 Avg: 52959.5
UDP_STREAM-20 : Avg: 30376.4 Avg: 29704
UDP_STREAM-40 : Avg: 15685.8 Avg: 15266.5
UDP_STREAM-80 : Avg: 8415.13 Avg: 7388.97
(wins and losses)
sysbench:
NO_WA_WEIGHT WA_WEIGHT
sysbench-mysql-2 : 2135.17 per sec. 2142.51 per sec.
sysbench-mysql-5 : 4809.68 per sec. 4800.19 per sec.
sysbench-mysql-10 : 9158.59 per sec. 9157.05 per sec.
sysbench-mysql-20 : 14570.70 per sec. 14543.55 per sec.
sysbench-mysql-40 : 22130.56 per sec. 22184.82 per sec.
sysbench-mysql-80 : 20995.56 per sec. 21904.18 per sec.
sysbench-psql-2 : 1679.58 per sec. 1705.06 per sec.
sysbench-psql-5 : 3797.69 per sec. 3879.93 per sec.
sysbench-psql-10 : 7253.22 per sec. 7258.06 per sec.
sysbench-psql-20 : 11166.75 per sec. 11220.00 per sec.
sysbench-psql-40 : 17277.28 per sec. 17359.78 per sec.
sysbench-psql-80 : 17112.44 per sec. 17221.16 per sec.
(increase on the top end)
tbench:
NO_WA_WEIGHT
Throughput 685.211 MB/sec 2 clients 2 procs max_latency=0.123 ms
Throughput 1596.64 MB/sec 5 clients 5 procs max_latency=0.119 ms
Throughput 2985.47 MB/sec 10 clients 10 procs max_latency=0.262 ms
Throughput 4521.15 MB/sec 20 clients 20 procs max_latency=0.506 ms
Throughput 9438.1 MB/sec 40 clients 40 procs max_latency=2.052 ms
Throughput 8210.5 MB/sec 80 clients 80 procs max_latency=8.310 ms
WA_WEIGHT
Throughput 697.292 MB/sec 2 clients 2 procs max_latency=0.127 ms
Throughput 1596.48 MB/sec 5 clients 5 procs max_latency=0.080 ms
Throughput 2975.22 MB/sec 10 clients 10 procs max_latency=0.254 ms
Throughput 4575.14 MB/sec 20 clients 20 procs max_latency=0.502 ms
Throughput 9468.65 MB/sec 40 clients 40 procs max_latency=2.069 ms
Throughput 8631.73 MB/sec 80 clients 80 procs max_latency=8.605 ms
(increase on the top end)
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Eric reported a sysbench regression against commit:
3fed382b46 ("sched/numa: Implement NUMA node level wake_affine()")
Similarly, Rik was looking at the NAS-lu.C benchmark, which regressed
against his v3.10 enterprise kernel.
PRE (current tip/master):
ivb-ep sysbench:
2: [30 secs] transactions: 64110 (2136.94 per sec.)
5: [30 secs] transactions: 143644 (4787.99 per sec.)
10: [30 secs] transactions: 274298 (9142.93 per sec.)
20: [30 secs] transactions: 418683 (13955.45 per sec.)
40: [30 secs] transactions: 320731 (10690.15 per sec.)
80: [30 secs] transactions: 355096 (11834.28 per sec.)
hsw-ex NAS:
OMP_PROC_BIND/lu.C.x_threads_144_run_1.log: Time in seconds = 18.01
OMP_PROC_BIND/lu.C.x_threads_144_run_2.log: Time in seconds = 17.89
OMP_PROC_BIND/lu.C.x_threads_144_run_3.log: Time in seconds = 17.93
lu.C.x_threads_144_run_1.log: Time in seconds = 434.68
lu.C.x_threads_144_run_2.log: Time in seconds = 405.36
lu.C.x_threads_144_run_3.log: Time in seconds = 433.83
POST (+patch):
ivb-ep sysbench:
2: [30 secs] transactions: 64494 (2149.75 per sec.)
5: [30 secs] transactions: 145114 (4836.99 per sec.)
10: [30 secs] transactions: 278311 (9276.69 per sec.)
20: [30 secs] transactions: 437169 (14571.60 per sec.)
40: [30 secs] transactions: 669837 (22326.73 per sec.)
80: [30 secs] transactions: 631739 (21055.88 per sec.)
hsw-ex NAS:
lu.C.x_threads_144_run_1.log: Time in seconds = 23.36
lu.C.x_threads_144_run_2.log: Time in seconds = 22.96
lu.C.x_threads_144_run_3.log: Time in seconds = 22.52
This patch takes out all the shiny wake_affine() stuff and goes back to
utter basics. Between the two CPUs involved with the wakeup (the CPU
doing the wakeup and the CPU we ran on previously) pick the CPU we can
run on _now_.
This restores much of the regressions against the older kernels,
but leaves some ground in the overloaded case. The default-enabled
WA_WEIGHT (which will be introduced in the next patch) is an attempt
to address the overloaded situation.
Reported-by: Eric Farman <farman@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Rosato <mjrosato@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: jinpuwang@gmail.com
Cc: vcaputo@pengaru.com
Fixes: 3fed382b46 ("sched/numa: Implement NUMA node level wake_affine()")
Signed-off-by: Ingo Molnar <mingo@kernel.org>
I had a wee bit of trouble recalling how the calc_group_runnable()
stuff worked.. add hopefully better comments.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Our runnable_weight currently looks like this
runnable_weight = shares * runnable_load_avg / load_avg
The goal is to scale the runnable weight for the group based on its runnable to
load_avg ratio. The problem with this is it biases us towards tasks that never
go to sleep. Tasks that go to sleep are going to have their runnable_load_avg
decayed pretty hard, which will drastically reduce the runnable weight of groups
with interactive tasks. To solve this imbalance we tweak this slightly, so in
the ideal case it is still the above, but in the interactive case it is
runnable_weight = shares * runnable_weight / load_weight
which will make the weight distribution fairer between interactive and
non-interactive groups.
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: kernel-team@fb.com
Cc: linux-kernel@vger.kernel.org
Cc: riel@redhat.com
Cc: tj@kernel.org
Link: http://lkml.kernel.org/r/1501773219-18774-2-git-send-email-jbacik@fb.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The problem with the overestimate is that it will subtract too big a
value from the load_sum, thereby pushing it down further than it ought
to go. Since runnable_load_avg is not subject to a similar 'force',
this results in the occasional 'runnable_load > load' situation.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The PELT _sum values are a saw-tooth function, dropping on the decay
edge and then growing back up again during the window.
When these window-edges are not aligned between cfs_rq and se, we can
have the situation where, for example, on dequeue, the se decays
first.
Its _sum values will be small(er), while the cfs_rq _sum values will
still be on their way up. Because of this, the subtraction:
cfs_rq->avg._sum -= se->avg._sum will result in a positive value. This
will then, once the cfs_rq reaches an edge, translate into its _avg
value jumping up.
This is especially visible with the runnable_load bits, since they get
added/subtracted a lot.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Vincent wondered why his self migrating task had a roughly 50% dip in
load_avg when landing on the new CPU. This is because we uncondionally
take the asynchronous detatch_entity route, which can lead to the
attach on the new CPU still seeing the old CPU's contribution to
tg->load_avg, effectively halving the new CPU's shares.
While in general this is something we have to live with, there is the
special case of runnable migration where we can do better.
Tested-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The load balancer uses runnable_load_avg as load indicator. For
!cgroup this is:
runnable_load_avg = \Sum se->avg.load_avg ; where se->on_rq
That is, a direct sum of all runnable tasks on that runqueue. As
opposed to load_avg, which is a sum of all tasks on the runqueue,
which includes a blocked component.
However, in the cgroup case, this comes apart since the group entities
are always runnable, even if most of their constituent entities are
blocked.
Therefore introduce a runnable_weight which for task entities is the
same as the regular weight, but for group entities is a fraction of
the entity weight and represents the runnable part of the group
runqueue.
Then propagate this load through the PELT hierarchy to arrive at an
effective runnable load avgerage -- which we should not confuse with
the canonical runnable load average.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When an entity migrates in (or out) of a runqueue, we need to add (or
remove) its contribution from the entire PELT hierarchy, because even
non-runnable entities are included in the load average sums.
In order to do this we have some propagation logic that updates the
PELT tree, however the way it 'propagates' the runnable (or load)
change is (more or less):
tg->weight * grq->avg.load_avg
ge->avg.load_avg = ------------------------------
tg->load_avg
But that is the expression for ge->weight, and per the definition of
load_avg:
ge->avg.load_avg := ge->weight * ge->avg.runnable_avg
That destroys the runnable_avg (by setting it to 1) we wanted to
propagate.
Instead directly propagate runnable_sum.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Since on wakeup migration we don't hold the rq->lock for the old CPU
we cannot update its state. Instead we add the removed 'load' to an
atomic variable and have the next update on that CPU collect and
process it.
Currently we have 2 atomic variables; which already have the issue
that they can be read out-of-sync. Also, two atomic ops on a single
cacheline is already more expensive than an uncontended lock.
Since we want to add more, convert the thing over to an explicit
cacheline with a lock in.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Now that we directly change load_avg and propagate that change into
the sums, sys_nice() and co should do the same, otherwise its possible
to confuse load accounting when we migrate near the weight change.
Fixes-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
[ Added changelog, fixed the call condition. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Link: http://lkml.kernel.org/r/20170517095045.GA8420@linaro.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When a (group) entity changes it's weight we should instantly change
its load_avg and propagate that change into the sums it is part of.
Because we use these values to predict future behaviour and are not
interested in its historical value.
Without this change, the change in load would need to propagate
through the average, by which time it could again have changed etc..
always chasing itself.
With this change, the cfs_rq load_avg sum will more accurately reflect
the current runnable and expected return of blocked load.
Reported-by: Paul Turner <pjt@google.com>
[josef: compile fix !SMP || !FAIR_GROUP]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Analogous to the existing {en,de}queue_runnable_load_avg() add helpers
for {en,de}queue_load_avg(). More users will follow.
Includes some code movement to avoid fwd declarations.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Move the entity migrate handling from enqueue_entity_load_avg() to
update_load_avg(). This has two benefits:
- {en,de}queue_entity_load_avg() will become purely about managing
runnable_load
- we can avoid a double update_tg_load_avg() and reduce pressure on
the global tg->shares cacheline
The reason we do this is so that we can change update_cfs_shares() to
change both weight and (future) runnable_weight. For this to work we
need to have the cfs_rq averages up-to-date (which means having done
the attach), but we need the cfs_rq->avg.runnable_avg to not yet
include the se's contribution (since se->on_rq == 0).
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Most call sites of update_load_avg() already have cfs_rq_of(se)
available, pass it down instead of recomputing it.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Remove the load from the load_sum for sched_entities, basically
turning load_sum into runnable_sum. This prepares for better
reweighting of group entities.
Since we now have different rules for computing load_avg, split
___update_load_avg() into two parts, ___update_load_sum() and
___update_load_avg().
So for se:
___update_load_sum(.weight = 1)
___upate_load_avg(.weight = se->load.weight)
and for cfs_rq:
___update_load_sum(.weight = cfs_rq->load.weight)
___upate_load_avg(.weight = 1)
Since the primary consumable is load_avg, most things will not be
affected. Only those few sites that initialize/modify load_sum need
attention.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Vincent reported that when running in a cgroup, his root
cfs_rq->avg.load_avg dropped to 0 on task idle.
This is because reweight_entity() will now immediately propagate the
weight change of the group entity to its cfs_rq, and as it happens,
our approxmation (5) for calc_cfs_shares() results in 0 when the group
is idle.
Avoid this by using the correct (3) as a lower bound on (5). This way
the empty cgroup will slowly decay instead of instantly drop to 0.
Reported-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Explain the magic equation in calc_cfs_shares() a bit better.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
For consistencies sake, we should have only a single reading of tg->shares.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Markus reported that tasks in TASK_IDLE state are reported by SysRq-W,
which results in undesirable clutter.
Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Now that we have added breaks in the wait queue scan and allow bookmark
on scan position, we put this logic in the wake_up_page_bit function.
We can have very long page wait list in large system where multiple
pages share the same wait list. We break the wake up walk here to allow
other cpus a chance to access the list, and not to disable the interrupts
when traversing the list for too long. This reduces the interrupt and
rescheduling latency, and excessive page wait queue lock hold time.
[ v2: Remove bookmark_wake_function ]
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We encountered workloads that have very long wake up list on large
systems. A waker takes a long time to traverse the entire wake list and
execute all the wake functions.
We saw page wait list that are up to 3700+ entries long in tests of
large 4 and 8 socket systems. It took 0.8 sec to traverse such list
during wake up. Any other CPU that contends for the list spin lock will
spin for a long time. It is a result of the numa balancing migration of
hot pages that are shared by many threads.
Multiple CPUs waking are queued up behind the lock, and the last one
queued has to wait until all CPUs did all the wakeups.
The page wait list is traversed with interrupt disabled, which caused
various problems. This was the original cause that triggered the NMI
watch dog timer in: https://patchwork.kernel.org/patch/9800303/ . Only
extending the NMI watch dog timer there helped.
This patch bookmarks the waker's scan position in wake list and break
the wake up walk, to allow access to the list before the waker resume
its walk down the rest of the wait list. It lowers the interrupt and
rescheduling latency.
This patch also provides a performance boost when combined with the next
patch to break up page wakeup list walk. We saw 22% improvement in the
will-it-scale file pread2 test on a Xeon Phi system running 256 threads.
[ v2: Merged in Linus' changes to remove the bookmark_wake_function, and
simply access to flags. ]
Reported-by: Kan Liang <kan.liang@intel.com>
Tested-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull scheduler fixes from Ingo Molnar:
"Three CPU hotplug related fixes and a debugging improvement"
* 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/debug: Add debugfs knob for "sched_debug"
sched/core: WARN() when migrating to an offline CPU
sched/fair: Plug hole between hotplug and active_load_balance()
sched/fair: Avoid newidle balance for !active CPUs
I'm forever late for editing my kernel cmdline, add a runtime knob to
disable the "sched_debug" thing.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20170907150614.142924283@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Migrating tasks to offline CPUs is a pretty big fail, warn about it.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20170907150614.094206976@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The load balancer applies cpu_active_mask to whatever sched_domains it
finds, however in the case of active_balance there is a hole between
setting rq->{active_balance,push_cpu} and running the stop_machine
work doing the actual migration.
The @push_cpu can go offline in this window, which would result in us
moving a task onto a dead cpu, which is a fairly bad thing.
Double check the active mask before the stop work does the migration.
CPU0 CPU1
<SoftIRQ>
stop_machine(takedown_cpu)
load_balance() cpu_stopper_thread()
... work = multi_cpu_stop
stop_one_cpu_nowait( /* wait for CPU0 */
.func = active_load_balance_cpu_stop
);
</SoftIRQ>
cpu_stopper_thread()
work = multi_cpu_stop
/* sync with CPU1 */
take_cpu_down()
<idle>
play_dead();
work = active_load_balance_cpu_stop
set_task_cpu(p, CPU1); /* oops!! */
Reported-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170907150614.044460912@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
On CPU hot unplug, when parking the last kthread we'll try and
schedule into idle to kill the CPU. This last schedule can (and does)
trigger newidle balance because at this point the sched domains are
still up because of commit:
77d1dfda0e ("sched/topology, cpuset: Avoid spurious/wrong domain rebuilds")
Obviously pulling tasks to an already offline CPU is a bad idea, and
all balancing operations _should_ be subject to cpu_active_mask, make
it so.
Reported-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Fixes: 77d1dfda0e ("sched/topology, cpuset: Avoid spurious/wrong domain rebuilds")
Link: http://lkml.kernel.org/r/20170907150613.994135806@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Work around kernel-doc warning ('*' in Sphinx doc means "emphasis"):
../kernel/sched/fair.c:7584: WARNING: Inline emphasis start-string without end-string.
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/f18b30f9-6251-6d86-9d44-16501e386891@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
First, number of CPUs can't be negative number.
Second, different signnnedness leads to suboptimal code in the following
cases:
1)
kmalloc(nr_cpu_ids * sizeof(X));
"int" has to be sign extended to size_t.
2)
while (loff_t *pos < nr_cpu_ids)
MOVSXD is 1 byte longed than the same MOV.
Other cases exist as well. Basically compiler is told that nr_cpu_ids
can't be negative which can't be deduced if it is "int".
Code savings on allyesconfig kernel: -3KB
add/remove: 0/0 grow/shrink: 25/264 up/down: 261/-3631 (-3370)
function old new delta
coretemp_cpu_online 450 512 +62
rcu_init_one 1234 1272 +38
pci_device_probe 374 399 +25
...
pgdat_reclaimable_pages 628 556 -72
select_fallback_rq 446 369 -77
task_numa_find_cpu 1923 1807 -116
Link: http://lkml.kernel.org/r/20170819114959.GA30580@avx2
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
