Skip to content

cloud-kernel
cloud-kernel

openanolis / cloud-kernel
1 年多 前同步成功

通知 161
Star 36
Fork 7
cloud-kernel
cloud-kernel
切换分支/标签
  1. 25 5月, 2018 1 次提交
  3. 14 5月, 2018 2 次提交
    • R
      sched/core: Distinguish between idle_cpu() calls based on desired effect,... · 943d355d
      由 Rohit Jain 提交于 
      sched/core: Distinguish between idle_cpu() calls based on desired effect, introduce available_idle_cpu()

In the following commit:

  247f2f6f ("sched/core: Don't schedule threads on pre-empted vCPUs")

... we distinguish between idle_cpu() when the vCPU is not running for
scheduling threads.

However, the idle_cpu() function is used in other places for
actually checking whether the state of the CPU is idle or not.

Hence split the use of that function based on the desired return value,
by introducing the available_idle_cpu() function.

This fixes a (slight) regression in that initial vCPU commit, because
some code paths (like the load-balancer) don't care and shouldn't care
if the vCPU is preempted or not, they just want to know if there's any
tasks on the CPU.
Signed-off-by: NRohit Jain <rohit.k.jain@oracle.com>
Signed-off-by: NPeter 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: dhaval.giani@oracle.com
Cc: linux-kernel@vger.kernel.org
Cc: matt@codeblueprint.co.uk
Cc: steven.sistare@oracle.com
Cc: subhra.mazumdar@oracle.com
Link: http://lkml.kernel.org/r/1525883988-10356-1-git-send-email-rohit.k.jain@oracle.comSigned-off-by: NIngo Molnar <mingo@kernel.org>
      943d355d
    • M
      sched/numa: Stagger NUMA balancing scan periods for new threads · 13784475
      由 Mel Gorman 提交于 
      Threads share an address space and each can change the protections of the
same address space to trap NUMA faults. This is redundant and potentially
counter-productive as any thread doing the update will suffice. Potentially
only one thread is required but that thread may be idle or it may not have
any locality concerns and pick an unsuitable scan rate.

This patch uses independent scan period but they are staggered based on
the number of address space users when the thread is created.  The intent
is that threads will avoid scanning at the same time and have a chance
to adapt their scan rate later if necessary. This reduces the total scan
activity early in the lifetime of the threads.

The different in headline performance across a range of machines and
workloads is marginal but the system CPU usage is reduced as well as overall
scan activity.  The following is the time reported by NAS Parallel Benchmark
using unbound openmp threads and a D size class:

			      4.17.0-rc1             4.17.0-rc1
				 vanilla           stagger-v1r1
	Time bt.D      442.77 (   0.00%)      419.70 (   5.21%)
	Time cg.D      171.90 (   0.00%)      180.85 (  -5.21%)
	Time ep.D       33.10 (   0.00%)       32.90 (   0.60%)
	Time is.D        9.59 (   0.00%)        9.42 (   1.77%)
	Time lu.D      306.75 (   0.00%)      304.65 (   0.68%)
	Time mg.D       54.56 (   0.00%)       52.38 (   4.00%)
	Time sp.D     1020.03 (   0.00%)      903.77 (  11.40%)
	Time ua.D      400.58 (   0.00%)      386.49 (   3.52%)

Note it's not a universal win but we have no prior knowledge of which
thread matters but the number of threads created often exceeds the size
of the node when the threads are not bound. However, there is a reducation
of overall system CPU usage:

				    4.17.0-rc1             4.17.0-rc1
				       vanilla           stagger-v1r1
	sys-time-bt.D         48.78 (   0.00%)       48.22 (   1.15%)
	sys-time-cg.D         25.31 (   0.00%)       26.63 (  -5.22%)
	sys-time-ep.D          1.65 (   0.00%)        0.62 (  62.42%)
	sys-time-is.D         40.05 (   0.00%)       24.45 (  38.95%)
	sys-time-lu.D         37.55 (   0.00%)       29.02 (  22.72%)
	sys-time-mg.D         47.52 (   0.00%)       34.92 (  26.52%)
	sys-time-sp.D        119.01 (   0.00%)      109.05 (   8.37%)
	sys-time-ua.D         51.52 (   0.00%)       45.13 (  12.40%)

NUMA scan activity is also reduced:

	NUMA alloc local               1042828     1342670
	NUMA base PTE updates        140481138    93577468
	NUMA huge PMD updates           272171      180766
	NUMA page range updates      279832690   186129660
	NUMA hint faults               1395972     1193897
	NUMA hint local faults          877925      855053
	NUMA hint local percent             62          71
	NUMA pages migrated           12057909     9158023

Similar observations are made for other thread-intensive workloads. System
CPU usage is lower even though the headline gains in performance tend to be
small. For example, specjbb 2005 shows almost no difference in performance
but scan activity is reduced by a third on a 4-socket box. I didn't find
a workload (thread intensive or otherwise) that suffered badly.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Link: http://lkml.kernel.org/r/20180504154109.mvrha2qo5wdl65vr@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      13784475
  5. 12 5月, 2018 1 次提交
    • M
      Revert "sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine()" · 789ba280
      由 Mel Gorman 提交于 
      This reverts commit 7347fc87.

Srikar Dronamra pointed out that while the commit in question did show
a performance improvement on ppc64, it did so at the cost of disabling
active CPU migration by automatic NUMA balancing which was not the intent.
The issue was that a serious flaw in the logic failed to ever active balance
if SD_WAKE_AFFINE was disabled on scheduler domains. Even when it's enabled,
the logic is still bizarre and against the original intent.

Investigation showed that fixing the patch in either the way he suggested,
using the correct comparison for jiffies values or introducing a new
numa_migrate_deferred variable in task_struct all perform similarly to a
revert with a mix of gains and losses depending on the workload, machine
and socket count.

The original intent of the commit was to handle a problem whereby
wake_affine, idle balancing and automatic NUMA balancing disagree on the
appropriate placement for a task. This was particularly true for cases where
a single task was a massive waker of tasks but where wake_wide logic did
not apply.  This was particularly noticeable when a futex (a barrier) woke
all worker threads and tried pulling the wakees to the waker nodes. In that
specific case, it could be handled by tuning MPI or openMP appropriately,
but the behavior is not illogical and was worth attempting to fix. However,
the approach was wrong. Given that we're at rc4 and a fix is not obvious,
it's better to play safe, revert this commit and retry later.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: NSrikar Dronamraju <srikar@linux.vnet.ibm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: efault@gmx.de
Cc: ggherdovich@suse.cz
Cc: hpa@zytor.com
Cc: matt@codeblueprint.co.uk
Cc: mpe@ellerman.id.au
Link: http://lkml.kernel.org/r/20180509163115.6fnnyeg4vdm2ct4v@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      789ba280
  7. 04 5月, 2018 2 次提交
  9. 03 5月, 2018 1 次提交
  11. 05 4月, 2018 1 次提交
  13. 20 3月, 2018 3 次提交
    • P
      sched/fair: Update util_est only on util_avg updates · d519329f
      由 Patrick Bellasi 提交于 
      The estimated utilization of a task is currently updated every time the
task is dequeued. However, to keep overheads under control, PELT signals
are effectively updated at maximum once every 1ms.

Thus, for really short running tasks, it can happen that their util_avg
value has not been updates since their last enqueue.  If such tasks are
also frequently running tasks (e.g. the kind of workload generated by
hackbench) it can also happen that their util_avg is updated only every
few activations.

This means that updating util_est at every dequeue potentially introduces
not necessary overheads and it's also conceptually wrong if the util_avg
signal has never been updated during a task activation.

Let's introduce a throttling mechanism on task's util_est updates
to sync them with util_avg updates. To make the solution memory
efficient, both in terms of space and load/store operations, we encode a
synchronization flag into the LSB of util_est.enqueued.
This makes util_est an even values only metric, which is still
considered good enough for its purpose.
The synchronization bit is (re)set by __update_load_avg_se() once the
PELT signal of a task has been updated during its last activation.

Such a throttling mechanism allows to keep under control util_est
overheads in the wakeup hot path, thus making it a suitable mechanism
which can be enabled also on high-intensity workload systems.
Thus, this now switches on by default the estimation utilization
scheduler feature.
Suggested-by: NChris Redpath <chris.redpath@arm.com>
Signed-off-by: NPatrick Bellasi <patrick.bellasi@arm.com>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Paul Turner <pjt@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com>
Cc: Steve Muckle <smuckle@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Todd Kjos <tkjos@android.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Viresh Kumar <viresh.kumar@linaro.org>
Link: http://lkml.kernel.org/r/20180309095245.11071-5-patrick.bellasi@arm.comSigned-off-by: NIngo Molnar <mingo@kernel.org>
      d519329f
    • P
      sched/fair: Use util_est in LB and WU paths · f9be3e59
      由 Patrick Bellasi 提交于 
      When the scheduler looks at the CPU utilization, the current PELT value
for a CPU is returned straight away. In certain scenarios this can have
undesired side effects on task placement.

For example, since the task utilization is decayed at wakeup time, when
a long sleeping big task is enqueued it does not add immediately a
significant contribution to the target CPU.
As a result we generate a race condition where other tasks can be placed
on the same CPU while it is still considered relatively empty.

In order to reduce this kind of race conditions, this patch introduces the
required support to integrate the usage of the CPU's estimated utilization
in the wakeup path, via cpu_util_wake(), as well as in the load-balance
path, via cpu_util() which is used by update_sg_lb_stats().

The estimated utilization of a CPU is defined to be the maximum between
its PELT's utilization and the sum of the estimated utilization (at
previous dequeue time) of all the tasks currently RUNNABLE on that CPU.
This allows to properly represent the spare capacity of a CPU which, for
example, has just got a big task running since a long sleep period.
Signed-off-by: NPatrick Bellasi <patrick.bellasi@arm.com>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: NDietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Paul Turner <pjt@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com>
Cc: Steve Muckle <smuckle@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Todd Kjos <tkjos@android.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Viresh Kumar <viresh.kumar@linaro.org>
Link: http://lkml.kernel.org/r/20180309095245.11071-3-patrick.bellasi@arm.comSigned-off-by: NIngo Molnar <mingo@kernel.org>
      f9be3e59
    • P
      sched/fair: Add util_est on top of PELT · 7f65ea42
      由 Patrick Bellasi 提交于 
      The util_avg signal computed by PELT is too variable for some use-cases.
For example, a big task waking up after a long sleep period will have its
utilization almost completely decayed. This introduces some latency before
schedutil will be able to pick the best frequency to run a task.

The same issue can affect task placement. Indeed, since the task
utilization is already decayed at wakeup, when the task is enqueued in a
CPU, this can result in a CPU running a big task as being temporarily
represented as being almost empty. This leads to a race condition where
other tasks can be potentially allocated on a CPU which just started to run
a big task which slept for a relatively long period.

Moreover, the PELT utilization of a task can be updated every [ms], thus
making it a continuously changing value for certain longer running
tasks. This means that the instantaneous PELT utilization of a RUNNING
task is not really meaningful to properly support scheduler decisions.

For all these reasons, a more stable signal can do a better job of
representing the expected/estimated utilization of a task/cfs_rq.
Such a signal can be easily created on top of PELT by still using it as
an estimator which produces values to be aggregated on meaningful
events.

This patch adds a simple implementation of util_est, a new signal built on
top of PELT's util_avg where:

    util_est(task) = max(task::util_avg, f(task::util_avg@dequeue))

This allows to remember how big a task has been reported by PELT in its
previous activations via f(task::util_avg@dequeue), which is the new
_task_util_est(struct task_struct*) function added by this patch.

If a task should change its behavior and it runs longer in a new
activation, after a certain time its util_est will just track the
original PELT signal (i.e. task::util_avg).

The estimated utilization of cfs_rq is defined only for root ones.
That's because the only sensible consumer of this signal are the
scheduler and schedutil when looking for the overall CPU utilization
due to FAIR tasks.

For this reason, the estimated utilization of a root cfs_rq is simply
defined as:

    util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued)

where:

    cfs_rq::util_est::enqueued = sum(_task_util_est(task))
                                 for each RUNNABLE task on that root cfs_rq

It's worth noting that the estimated utilization is tracked only for
objects of interests, specifically:

 - Tasks: to better support tasks placement decisions
 - root cfs_rqs: to better support both tasks placement decisions as
                 well as frequencies selection
Signed-off-by: NPatrick Bellasi <patrick.bellasi@arm.com>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: NDietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Morten Rasmussen <morten.rasmussen@arm.com>
Cc: Paul Turner <pjt@google.com>
Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com>
Cc: Steve Muckle <smuckle@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Todd Kjos <tkjos@android.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Viresh Kumar <viresh.kumar@linaro.org>
Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.comSigned-off-by: NIngo Molnar <mingo@kernel.org>
      7f65ea42
  15. 09 3月, 2018 15 次提交
  17. 04 3月, 2018 1 次提交
    • I
      sched/headers: Simplify and clean up header usage in the scheduler · 325ea10c
      由 Ingo Molnar 提交于 
      Do the following cleanups and simplifications:

 - sched/sched.h already includes <asm/paravirt.h>, so no need to
   include it in sched/core.c again.

 - order the <linux/sched/*.h> headers alphabetically

 - add all <linux/sched/*.h> headers to kernel/sched/sched.h

 - remove all unnecessary includes from the .c files that
   are already included in kernel/sched/sched.h.

Finally, make all scheduler .c files use a single common header:

  #include "sched.h"

... which now contains a union of the relied upon headers.

This makes the various .c files easier to read and easier to handle.

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: NIngo Molnar <mingo@kernel.org>
      325ea10c
  19. 03 3月, 2018 1 次提交
    • I
      sched: Clean up and harmonize the coding style of the scheduler code base · 97fb7a0a
      由 Ingo Molnar 提交于 
      A good number of small style inconsistencies have accumulated
in the scheduler core, so do a pass over them to harmonize
all these details:

 - fix speling in comments,

 - use curly braces for multi-line statements,

 - remove unnecessary parentheses from integer literals,

 - capitalize consistently,

 - remove stray newlines,

 - add comments where necessary,

 - remove invalid/unnecessary comments,

 - align structure definitions and other data types vertically,

 - add missing newlines for increased readability,

 - fix vertical tabulation where it's misaligned,

 - harmonize preprocessor conditional block labeling
   and vertical alignment,

 - remove line-breaks where they uglify the code,

 - add newline after local variable definitions,

No change in functionality:

  md5:
     1191fa0a890cfa8132156d2959d7e9e2  built-in.o.before.asm
     1191fa0a890cfa8132156d2959d7e9e2  built-in.o.after.asm

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: NIngo Molnar <mingo@kernel.org>
      97fb7a0a
  21. 21 2月, 2018 7 次提交
    • F
      sched/isolation: Offload residual 1Hz scheduler tick · d84b3131
      由 Frederic Weisbecker 提交于 
      When a CPU runs in full dynticks mode, a 1Hz tick remains in order to
keep the scheduler stats alive. However this residual tick is a burden
for bare metal tasks that can't stand any interruption at all, or want
to minimize them.

The usual boot parameters "nohz_full=" or "isolcpus=nohz" will now
outsource these scheduler ticks to the global workqueue so that a
housekeeping CPU handles those remotely. The sched_class::task_tick()
implementations have been audited and look safe to be called remotely
as the target runqueue and its current task are passed in parameter
and don't seem to be accessed locally.

Note that in the case of using isolcpus, it's still up to the user to
affine the global workqueues to the housekeeping CPUs through
/sys/devices/virtual/workqueue/cpumask or domains isolation
"isolcpus=nohz,domain".
Signed-off-by: NFrederic Weisbecker <frederic@kernel.org>
Reviewed-by: NThomas Gleixner <tglx@linutronix.de>
Acked-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Chris Metcalf <cmetcalf@mellanox.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Wanpeng Li <kernellwp@gmail.com>
Link: http://lkml.kernel.org/r/1519186649-3242-6-git-send-email-frederic@kernel.orgSigned-off-by: NIngo Molnar <mingo@kernel.org>
      d84b3131
    • M
      sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine() · 7347fc87
      由 Mel Gorman 提交于 
      If wake_affine() pulls a task to another node for any reason and the node is
no longer preferred then temporarily stop automatic NUMA balancing pulling
the task back. Otherwise, tasks with a strong waker/wakee relationship
may constantly fight automatic NUMA balancing over where a task should
be placed.

Once again netperf is interesting here. The performance barely changes
but automatic NUMA balancing is interesting:

 Hmean     send-64         354.67 (   0.00%)      352.15 (  -0.71%)
 Hmean     send-128        702.91 (   0.00%)      693.84 (  -1.29%)
 Hmean     send-256       1350.07 (   0.00%)     1344.19 (  -0.44%)
 Hmean     send-1024      5124.38 (   0.00%)     4941.24 (  -3.57%)
 Hmean     send-2048      9687.44 (   0.00%)     9624.45 (  -0.65%)
 Hmean     send-3312     14577.64 (   0.00%)    14514.35 (  -0.43%)
 Hmean     send-4096     16393.62 (   0.00%)    16488.30 (   0.58%)
 Hmean     send-8192     26877.26 (   0.00%)    26431.63 (  -1.66%)
 Hmean     send-16384    38683.43 (   0.00%)    38264.91 (  -1.08%)
 Hmean     recv-64         354.67 (   0.00%)      352.15 (  -0.71%)
 Hmean     recv-128        702.91 (   0.00%)      693.84 (  -1.29%)
 Hmean     recv-256       1350.07 (   0.00%)     1344.19 (  -0.44%)
 Hmean     recv-1024      5124.38 (   0.00%)     4941.24 (  -3.57%)
 Hmean     recv-2048      9687.43 (   0.00%)     9624.45 (  -0.65%)
 Hmean     recv-3312     14577.59 (   0.00%)    14514.35 (  -0.43%)
 Hmean     recv-4096     16393.55 (   0.00%)    16488.20 (   0.58%)
 Hmean     recv-8192     26876.96 (   0.00%)    26431.29 (  -1.66%)
 Hmean     recv-16384    38682.41 (   0.00%)    38263.94 (  -1.08%)

 NUMA alloc hit                 1465986     1423090
 NUMA alloc miss                      0           0
 NUMA interleave hit                  0           0
 NUMA alloc local               1465897     1423003
 NUMA base PTE updates             1473        1420
 NUMA huge PMD updates                0           0
 NUMA page range updates           1473        1420
 NUMA hint faults                  1383        1312
 NUMA hint local faults             451         124
 NUMA hint local percent             32           9

There is a slight degrading in performance but there are slightly fewer
NUMA faults. There is a large drop in the percentage of local faults but
the bulk of migrations for netperf are in small shared libraries so it's
reflecting the fact that automatic NUMA balancing has backed off. This is
a case where despite wake_affine() and automatic NUMA balancing fighting
for placement that there is a marginal benefit to rescheduling to local
data quickly. However, it should be noted that wake_affine() and automatic
NUMA balancing fighting each other constantly is undesirable.

However, the benefit in other cases is large. This is the result for NAS
with the D class sizing on a 4-socket machine:

 nas-mpi
                           4.15.0                 4.15.0
                     sdnuma-v1r23       delayretry-v1r23
 Time cg.D      557.00 (   0.00%)      431.82 (  22.47%)
 Time ep.D       77.83 (   0.00%)       79.01 (  -1.52%)
 Time is.D       26.46 (   0.00%)       26.64 (  -0.68%)
 Time lu.D      727.14 (   0.00%)      597.94 (  17.77%)
 Time mg.D      191.35 (   0.00%)      146.85 (  23.26%)

               4.15.0      4.15.0
         sdnuma-v1r23delayretry-v1r23
 User        75665.20    70413.30
 System      20321.59     8861.67
 Elapsed       766.13      634.92

 Minor Faults                  16528502     7127941
 Major Faults                      4553        5068
 NUMA alloc local               6963197     6749135
 NUMA base PTE updates        366409093   107491434
 NUMA huge PMD updates           687556      198880
 NUMA page range updates      718437765   209317994
 NUMA hint faults              13643410     4601187
 NUMA hint local faults         9212593     3063996
 NUMA hint local percent             67          66

Note the massive reduction in system CPU usage even though the percentage
of local faults is barely affected. There is a massive reduction in the
number of PTE updates showing that automatic NUMA balancing has backed off.
A critical observation is also that there is a massive reduction in minor
faults which is due to far fewer NUMA hinting faults being trapped.

There were questions on NAS OMP and how it behaved related to threads
being bound to CPUs. First, there are more gains than losses with this
patch applied and a reduction in system CPU usage:

nas-omp
                      4.16.0-rc1             4.16.0-rc1
                     sdnuma-v2r1        delayretry-v2r1
Time bt.D      436.71 (   0.00%)      430.05 (   1.53%)
Time cg.D      201.02 (   0.00%)      180.87 (  10.02%)
Time ep.D       32.84 (   0.00%)       32.68 (   0.49%)
Time is.D        9.63 (   0.00%)        9.64 (  -0.10%)
Time lu.D      331.20 (   0.00%)      304.80 (   7.97%)
Time mg.D       54.87 (   0.00%)       52.72 (   3.92%)
Time sp.D     1108.78 (   0.00%)      917.10 (  17.29%)
Time ua.D      378.81 (   0.00%)      398.83 (  -5.28%)

          4.16.0-rc1  4.16.0-rc1
         sdnuma-v2r1delayretry-v2r1
User       305633.08   296751.91
System        451.75      357.80
Elapsed      2595.73     2368.13

However, it does not close the gap between binding and being unbound. There
is negligible difference between the performance of the baseline and a
patched kernel when threads are bound so it is not presented here:

                      4.16.0-rc1             4.16.0-rc1
                 delayretry-bind     delayretry-unbound
Time bt.D      385.02 (   0.00%)      430.05 ( -11.70%)
Time cg.D      144.02 (   0.00%)      180.87 ( -25.59%)
Time ep.D       32.85 (   0.00%)       32.68 (   0.52%)
Time is.D       10.52 (   0.00%)        9.64 (   8.37%)
Time lu.D      285.31 (   0.00%)      304.80 (  -6.83%)
Time mg.D       43.21 (   0.00%)       52.72 ( -22.01%)
Time sp.D      820.24 (   0.00%)      917.10 ( -11.81%)
Time ua.D      337.09 (   0.00%)      398.83 ( -18.32%)

          4.16.0-rc1  4.16.0-rc1
        delayretry-binddelayretry-unbound
User       277731.25   296751.91
System        261.29      357.80
Elapsed      2100.55     2368.13

Unfortunately, while performance is improved by the patch, there is still
quite a long way to go before it's equivalent to hard binding.

Other workloads like hackbench, tbench, dbench and schbench are barely
affected. dbench shows a mix of gains and losses depending on the machine
although in general, the results are more stable.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Giovanni Gherdovich <ggherdovich@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180213133730.24064-7-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      7347fc87
    • M
      sched/fair: Consider SD_NUMA when selecting the most idle group to schedule on · 2c833627
      由 Mel Gorman 提交于 
      find_idlest_group() compares a local group with each other group to select
the one that is most idle. When comparing groups in different NUMA domains,
a very slight imbalance is enough to select a remote NUMA node even if the
runnable load on both groups is 0 or close to 0. This ignores the cost of
remote accesses entirely and is a problem when selecting the CPU for a
newly forked task to run on.  This is problematic when a forking server
is almost guaranteed to run on a remote node incurring numerous remote
accesses and potentially causing automatic NUMA balancing to try migrate
the task back or migrate the data to another node. Similar weirdness is
observed if a basic shell command pipes output to another as each process
in the pipeline is likely to start on different nodes and then get adjusted
later by wake_affine().

This patch adds imbalance to remote domains when considering whether to
select CPUs from remote domains. If the local domain is selected, imbalance
will still be used to try select a CPU from a lower scheduler domain's group
instead of stacking tasks on the same CPU.

A variety of workloads and machines were tested and as expected, there is no
difference on UMA. The difference on NUMA can be dramatic. This is a comparison
of elapsed times running the git regression test suite. It's fork-intensive with
short-lived processes:

                                  4.15.0                 4.15.0
                            noexit-v1r23           sdnuma-v1r23
 Elapsed min          1706.06 (   0.00%)     1435.94 (  15.83%)
 Elapsed mean         1709.53 (   0.00%)     1436.98 (  15.94%)
 Elapsed stddev          2.16 (   0.00%)        1.01 (  53.38%)
 Elapsed coeffvar        0.13 (   0.00%)        0.07 (  44.54%)
 Elapsed max          1711.59 (   0.00%)     1438.01 (  15.98%)

               4.15.0      4.15.0
         noexit-v1r23 sdnuma-v1r23
 User         5434.12     5188.41
 System       4878.77     3467.09
 Elapsed     10259.06     8624.21

That shows a considerable reduction in elapsed times. It's important to
note that automatic NUMA balancing does not affect this load as processes
are too short-lived.

There is also a noticable impact on hackbench such as this example using
processes and pipes:

 hackbench-process-pipes
                               4.15.0                 4.15.0
                         noexit-v1r23           sdnuma-v1r23
 Amean     1        1.0973 (   0.00%)      0.9393 (  14.40%)
 Amean     4        1.3427 (   0.00%)      1.3730 (  -2.26%)
 Amean     7        1.4233 (   0.00%)      1.6670 ( -17.12%)
 Amean     12       3.0250 (   0.00%)      3.3013 (  -9.13%)
 Amean     21       9.0860 (   0.00%)      9.5343 (  -4.93%)
 Amean     30      14.6547 (   0.00%)     13.2433 (   9.63%)
 Amean     48      22.5447 (   0.00%)     20.4303 (   9.38%)
 Amean     79      29.2010 (   0.00%)     26.7853 (   8.27%)
 Amean     110     36.7443 (   0.00%)     35.8453 (   2.45%)
 Amean     141     45.8533 (   0.00%)     42.6223 (   7.05%)
 Amean     172     55.1317 (   0.00%)     50.6473 (   8.13%)
 Amean     203     64.4420 (   0.00%)     58.3957 (   9.38%)
 Amean     234     73.2293 (   0.00%)     67.1047 (   8.36%)
 Amean     265     80.5220 (   0.00%)     75.7330 (   5.95%)
 Amean     296     88.7567 (   0.00%)     82.1533 (   7.44%)

It's not a universal win as there are occasions when spreading wide and
quickly is a benefit but it's more of a win than it is a loss. For other
workloads, there is little difference but netperf is interesting. Without
the patch, the server and client starts on different nodes but quickly get
migrated due to wake_affine. Hence, the difference is overall performance
is marginal but detectable:

                                      4.15.0                 4.15.0
                                noexit-v1r23           sdnuma-v1r23
 Hmean     send-64         349.09 (   0.00%)      354.67 (   1.60%)
 Hmean     send-128        699.16 (   0.00%)      702.91 (   0.54%)
 Hmean     send-256       1316.34 (   0.00%)     1350.07 (   2.56%)
 Hmean     send-1024      5063.99 (   0.00%)     5124.38 (   1.19%)
 Hmean     send-2048      9705.19 (   0.00%)     9687.44 (  -0.18%)
 Hmean     send-3312     14359.48 (   0.00%)    14577.64 (   1.52%)
 Hmean     send-4096     16324.20 (   0.00%)    16393.62 (   0.43%)
 Hmean     send-8192     26112.61 (   0.00%)    26877.26 (   2.93%)
 Hmean     send-16384    37208.44 (   0.00%)    38683.43 (   3.96%)
 Hmean     recv-64         349.09 (   0.00%)      354.67 (   1.60%)
 Hmean     recv-128        699.16 (   0.00%)      702.91 (   0.54%)
 Hmean     recv-256       1316.34 (   0.00%)     1350.07 (   2.56%)
 Hmean     recv-1024      5063.99 (   0.00%)     5124.38 (   1.19%)
 Hmean     recv-2048      9705.16 (   0.00%)     9687.43 (  -0.18%)
 Hmean     recv-3312     14359.42 (   0.00%)    14577.59 (   1.52%)
 Hmean     recv-4096     16323.98 (   0.00%)    16393.55 (   0.43%)
 Hmean     recv-8192     26111.85 (   0.00%)    26876.96 (   2.93%)
 Hmean     recv-16384    37206.99 (   0.00%)    38682.41 (   3.97%)

However, what is very interesting is how automatic NUMA balancing behaves.
Each netperf instance runs long enough for balancing to activate:

 NUMA base PTE updates             4620        1473
 NUMA huge PMD updates                0           0
 NUMA page range updates           4620        1473
 NUMA hint faults                  4301        1383
 NUMA hint local faults            1309         451
 NUMA hint local percent             30          32
 NUMA pages migrated               1335         491
 AutoNUMA cost                      21%          6%

There is an unfortunate number of remote faults although tracing indicated
that the vast majority are in shared libraries. However, the tendency to
start tasks on the same node if there is capacity means that there were
far fewer PTE updates and faults incurred overall.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Giovanni Gherdovich <ggherdovich@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180213133730.24064-6-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      2c833627
    • P
      sched/fair: Do not migrate due to a sync wakeup on exit · 24d0c1d6
      由 Peter Zijlstra 提交于 
      When a task exits, it notifies the parent that it has exited. This is a
sync wakeup and the exiting task may pull the parent towards the wakers
CPU. For simple workloads like using a shell, it was observed that the
shell is pulled across nodes by exiting processes. This is daft as the
parent may be long-lived and properly placed. This patch special cases a
sync wakeup on exit to avoid pulling tasks across nodes. Testing on a range
of workloads and machines showed very little differences in performance
although there was a small 3% boost on some machines running a shellscript
intensive workload (git regression test suite).
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Giovanni Gherdovich <ggherdovich@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180213133730.24064-5-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      24d0c1d6
    • M
      sched/fair: Do not migrate on wake_affine_weight() if weights are equal · 082f764a
      由 Mel Gorman 提交于 
      wake_affine_weight() will consider migrating a task to, or near, the current
CPU if there is a load imbalance. If the CPUs share LLC then either CPU
is valid as a search-for-idle-sibling target and equally appropriate for
stacking two tasks on one CPU if an idle sibling is unavailable. If they do
not share cache then a cross-node migration potentially impacts locality
so while they are equal from a CPU capacity point of view, they are not
equal in terms of memory locality. In either case, it's more appropriate
to migrate only if there is a difference in their effective load.

This patch modifies wake_affine_weight() to only consider migrating a task
if there is a load imbalance for normal wakeups but will allow potential
stacking if the loads are equal and it's a sync wakeup.

For the most part, the different in performance is marginal. For example,
on a 4-socket server running netperf UDP_STREAM on localhost the differences
are as follows:

                                      4.15.0                 4.15.0
                                       16rc0          noequal-v1r23
 Hmean     send-64         355.47 (   0.00%)      349.50 (  -1.68%)
 Hmean     send-128        697.98 (   0.00%)      693.35 (  -0.66%)
 Hmean     send-256       1328.02 (   0.00%)     1318.77 (  -0.70%)
 Hmean     send-1024      5051.83 (   0.00%)     5051.11 (  -0.01%)
 Hmean     send-2048      9637.02 (   0.00%)     9601.34 (  -0.37%)
 Hmean     send-3312     14355.37 (   0.00%)    14414.51 (   0.41%)
 Hmean     send-4096     16464.97 (   0.00%)    16301.37 (  -0.99%)
 Hmean     send-8192     26722.42 (   0.00%)    26428.95 (  -1.10%)
 Hmean     send-16384    38137.81 (   0.00%)    38046.11 (  -0.24%)
 Hmean     recv-64         355.47 (   0.00%)      349.50 (  -1.68%)
 Hmean     recv-128        697.98 (   0.00%)      693.35 (  -0.66%)
 Hmean     recv-256       1328.02 (   0.00%)     1318.77 (  -0.70%)
 Hmean     recv-1024      5051.83 (   0.00%)     5051.11 (  -0.01%)
 Hmean     recv-2048      9636.95 (   0.00%)     9601.30 (  -0.37%)
 Hmean     recv-3312     14355.32 (   0.00%)    14414.48 (   0.41%)
 Hmean     recv-4096     16464.74 (   0.00%)    16301.16 (  -0.99%)
 Hmean     recv-8192     26721.63 (   0.00%)    26428.17 (  -1.10%)
 Hmean     recv-16384    38136.00 (   0.00%)    38044.88 (  -0.24%)
 Stddev    send-64           7.30 (   0.00%)        4.75 (  34.96%)
 Stddev    send-128         15.15 (   0.00%)       22.38 ( -47.66%)
 Stddev    send-256         13.99 (   0.00%)       19.14 ( -36.81%)
 Stddev    send-1024       105.73 (   0.00%)       67.38 (  36.27%)
 Stddev    send-2048       294.57 (   0.00%)      223.88 (  24.00%)
 Stddev    send-3312       302.28 (   0.00%)      271.74 (  10.10%)
 Stddev    send-4096       195.92 (   0.00%)      121.10 (  38.19%)
 Stddev    send-8192       399.71 (   0.00%)      563.77 ( -41.04%)
 Stddev    send-16384     1163.47 (   0.00%)     1103.68 (   5.14%)
 Stddev    recv-64           7.30 (   0.00%)        4.75 (  34.96%)
 Stddev    recv-128         15.15 (   0.00%)       22.38 ( -47.66%)
 Stddev    recv-256         13.99 (   0.00%)       19.14 ( -36.81%)
 Stddev    recv-1024       105.73 (   0.00%)       67.38 (  36.27%)
 Stddev    recv-2048       294.59 (   0.00%)      223.89 (  24.00%)
 Stddev    recv-3312       302.24 (   0.00%)      271.75 (  10.09%)
 Stddev    recv-4096       196.03 (   0.00%)      121.14 (  38.20%)
 Stddev    recv-8192       399.86 (   0.00%)      563.65 ( -40.96%)
 Stddev    recv-16384     1163.79 (   0.00%)     1103.86 (   5.15%)

The difference in overall performance is marginal but note that most
measurements are less variable. There were similar observations for other
netperf comparisons. hackbench with sockets or threads with processes or
threads showed minor difference with some reduction of migration. tbench
showed only marginal differences that were within the noise. dbench,
regardless of filesystem, showed minor differences all of which are
within noise. Multiple machines, both UMA and NUMA were tested without
any regressions showing up.

The biggest risk with a patch like this is affecting wakeup latencies.
However, the schbench load from Facebook which is very sensitive to wakeup
latency showed a mixed result with mostly improvements in wakeup latency:

                                      4.15.0                 4.15.0
                                       16rc0          noequal-v1r23
 Lat 50.00th-qrtle-1        38.00 (   0.00%)       38.00 (   0.00%)
 Lat 75.00th-qrtle-1        49.00 (   0.00%)       41.00 (  16.33%)
 Lat 90.00th-qrtle-1        52.00 (   0.00%)       50.00 (   3.85%)
 Lat 95.00th-qrtle-1        54.00 (   0.00%)       51.00 (   5.56%)
 Lat 99.00th-qrtle-1        63.00 (   0.00%)       60.00 (   4.76%)
 Lat 99.50th-qrtle-1        66.00 (   0.00%)       61.00 (   7.58%)
 Lat 99.90th-qrtle-1        78.00 (   0.00%)       65.00 (  16.67%)
 Lat 50.00th-qrtle-2        38.00 (   0.00%)       38.00 (   0.00%)
 Lat 75.00th-qrtle-2        42.00 (   0.00%)       43.00 (  -2.38%)
 Lat 90.00th-qrtle-2        46.00 (   0.00%)       48.00 (  -4.35%)
 Lat 95.00th-qrtle-2        49.00 (   0.00%)       50.00 (  -2.04%)
 Lat 99.00th-qrtle-2        55.00 (   0.00%)       57.00 (  -3.64%)
 Lat 99.50th-qrtle-2        58.00 (   0.00%)       60.00 (  -3.45%)
 Lat 99.90th-qrtle-2        65.00 (   0.00%)       68.00 (  -4.62%)
 Lat 50.00th-qrtle-4        41.00 (   0.00%)       41.00 (   0.00%)
 Lat 75.00th-qrtle-4        45.00 (   0.00%)       46.00 (  -2.22%)
 Lat 90.00th-qrtle-4        50.00 (   0.00%)       50.00 (   0.00%)
 Lat 95.00th-qrtle-4        54.00 (   0.00%)       53.00 (   1.85%)
 Lat 99.00th-qrtle-4        61.00 (   0.00%)       61.00 (   0.00%)
 Lat 99.50th-qrtle-4        65.00 (   0.00%)       64.00 (   1.54%)
 Lat 99.90th-qrtle-4        76.00 (   0.00%)       82.00 (  -7.89%)
 Lat 50.00th-qrtle-8        48.00 (   0.00%)       46.00 (   4.17%)
 Lat 75.00th-qrtle-8        55.00 (   0.00%)       54.00 (   1.82%)
 Lat 90.00th-qrtle-8        60.00 (   0.00%)       59.00 (   1.67%)
 Lat 95.00th-qrtle-8        63.00 (   0.00%)       63.00 (   0.00%)
 Lat 99.00th-qrtle-8        71.00 (   0.00%)       69.00 (   2.82%)
 Lat 99.50th-qrtle-8        74.00 (   0.00%)       73.00 (   1.35%)
 Lat 99.90th-qrtle-8        98.00 (   0.00%)       90.00 (   8.16%)
 Lat 50.00th-qrtle-16       56.00 (   0.00%)       55.00 (   1.79%)
 Lat 75.00th-qrtle-16       68.00 (   0.00%)       67.00 (   1.47%)
 Lat 90.00th-qrtle-16       77.00 (   0.00%)       78.00 (  -1.30%)
 Lat 95.00th-qrtle-16       82.00 (   0.00%)       84.00 (  -2.44%)
 Lat 99.00th-qrtle-16       90.00 (   0.00%)       93.00 (  -3.33%)
 Lat 99.50th-qrtle-16       93.00 (   0.00%)       97.00 (  -4.30%)
 Lat 99.90th-qrtle-16      110.00 (   0.00%)      110.00 (   0.00%)
 Lat 50.00th-qrtle-32       68.00 (   0.00%)       62.00 (   8.82%)
 Lat 75.00th-qrtle-32       90.00 (   0.00%)       83.00 (   7.78%)
 Lat 90.00th-qrtle-32      110.00 (   0.00%)      100.00 (   9.09%)
 Lat 95.00th-qrtle-32      122.00 (   0.00%)      111.00 (   9.02%)
 Lat 99.00th-qrtle-32      145.00 (   0.00%)      133.00 (   8.28%)
 Lat 99.50th-qrtle-32      154.00 (   0.00%)      143.00 (   7.14%)
 Lat 99.90th-qrtle-32     2316.00 (   0.00%)      515.00 (  77.76%)
 Lat 50.00th-qrtle-35       69.00 (   0.00%)       72.00 (  -4.35%)
 Lat 75.00th-qrtle-35       92.00 (   0.00%)       95.00 (  -3.26%)
 Lat 90.00th-qrtle-35      111.00 (   0.00%)      114.00 (  -2.70%)
 Lat 95.00th-qrtle-35      122.00 (   0.00%)      124.00 (  -1.64%)
 Lat 99.00th-qrtle-35      142.00 (   0.00%)      144.00 (  -1.41%)
 Lat 99.50th-qrtle-35      150.00 (   0.00%)      154.00 (  -2.67%)
 Lat 99.90th-qrtle-35     6104.00 (   0.00%)     5640.00 (   7.60%)
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Giovanni Gherdovich <ggherdovich@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180213133730.24064-4-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      082f764a
    • M
      sched/fair: Defer calculation of 'prev_eff_load' in wake_affine_weight() until needed · eeb60398
      由 Mel Gorman 提交于 
      On sync wakeups, the previous CPU effective load may not be used so delay
the calculation until it's needed.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Giovanni Gherdovich <ggherdovich@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180213133730.24064-3-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      eeb60398
    • M
      sched/fair: Avoid an unnecessary lookup of current CPU ID during wake_affine · 7ebb66a1
      由 Mel Gorman 提交于 
      The only caller of wake_affine() knows the CPU ID. Pass it in instead of
rechecking it.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Giovanni Gherdovich <ggherdovich@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180213133730.24064-2-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      7ebb66a1
  23. 13 2月, 2018 1 次提交
  25. 06 2月, 2018 4 次提交
    • M
      sched/fair: Use a recently used CPU as an idle candidate and the basis for SIS · 32e839dd
      由 Mel Gorman 提交于 
      The select_idle_sibling() (SIS) rewrite in commit:

  10e2f1ac ("sched/core: Rewrite and improve select_idle_siblings()")

... replaced a domain iteration with a search that broadly speaking
does a wrapped walk of the scheduler domain sharing a last-level-cache.

While this had a number of improvements, one consequence is that two tasks
that share a waker/wakee relationship push each other around a socket. Even
though two tasks may be active, all cores are evenly used. This is great from
a search perspective and spreads a load across individual cores, but it has
adverse consequences for cpufreq. As each CPU has relatively low utilisation,
cpufreq may decide the utilisation is too low to used a higher P-state and
overall computation throughput suffers.

While individual cpufreq and cpuidle drivers may compensate by artifically
boosting P-state (at c0) or avoiding lower C-states (during idle), it does
not help if hardware-based cpufreq (e.g. HWP) is used.

This patch tracks a recently used CPU based on what CPU a task was running
on when it last was a waker a CPU it was recently using when a task is a
wakee. During SIS, the recently used CPU is used as a target if it's still
allowed by the task and is idle.

The benefit may be non-obvious so consider an example of two tasks
communicating back and forth. Task A may be an application doing IO where
task B is a kworker or kthread like journald. Task A may issue IO, wake
B and B wakes up A on completion.  With the existing scheme this may look
like the following (potentially different IDs if SMT is in use but similar
principal applies).

 A (cpu 0)	wake	B (wakes on cpu 1)
 B (cpu 1)	wake	A (wakes on cpu 2)
 A (cpu 2)	wake	B (wakes on cpu 3)
 etc.

A careful reader may wonder why CPU 0 was not idle when B wakes A the
first time and it's simply due to the fact that A can be rescheduled to
another CPU and the pattern is that prev == target when B tries to wakeup A
and the information about CPU 0 has been lost.

With this patch, the pattern is more likely to be:

 A (cpu 0)	wake	B (wakes on cpu 1)
 B (cpu 1)	wake	A (wakes on cpu 0)
 A (cpu 0)	wake	B (wakes on cpu 1)
 etc

i.e. two communicating casts are more likely to use just two cores instead
of all available cores sharing a LLC.

The most dramatic speedup was noticed on dbench using the XFS filesystem on
UMA as clients interact heavily with workqueues in that configuration. Note
that a similar speedup is not observed on ext4 as the wakeup pattern
is different:

                          4.15.0-rc9             4.15.0-rc9
                           waprev-v1        biasancestor-v1
 Hmean      1      287.54 (   0.00%)      817.01 ( 184.14%)
 Hmean      2     1268.12 (   0.00%)     1781.24 (  40.46%)
 Hmean      4     1739.68 (   0.00%)     1594.47 (  -8.35%)
 Hmean      8     2464.12 (   0.00%)     2479.56 (   0.63%)
 Hmean     64     1455.57 (   0.00%)     1434.68 (  -1.44%)

The results can be less dramatic on NUMA where automatic balancing interferes
with the test. It's also known that network benchmarks running on localhost
also benefit quite a bit from this patch (roughly 10% on netperf RR for UDP
and TCP depending on the machine). Hackbench also seens small improvements
(6-11% depending on machine and thread count). The facebook schbench was also
tested but in most cases showed little or no different to wakeup latencies.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180130104555.4125-5-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      32e839dd
    • M
      sched/fair: Do not migrate if the prev_cpu is idle · 806486c3
      由 Mel Gorman 提交于 
      wake_affine_idle() prefers to move a task to the current CPU if the
wakeup is due to an interrupt. The expectation is that the interrupt
data is cache hot and relevant to the waking task as well as avoiding
a search. However, there is no way to determine if there was cache hot
data on the previous CPU that may exceed the interrupt data. Furthermore,
round-robin delivery of interrupts can migrate tasks around a socket where
each CPU is under-utilised.  This can interact badly with cpufreq which
makes decisions based on per-cpu data. It has been observed on machines
with HWP that p-states are not boosted to their maximum levels even though
the workload is latency and throughput sensitive.

This patch uses the previous CPU for the task if it's idle and cache-affine
with the current CPU even if the current CPU is idle due to the wakup
being related to the interrupt. This reduces migrations at the cost of
the interrupt data not being cache hot when the task wakes.

A variety of workloads were tested on various machines and no adverse
impact was noticed that was outside noise. dbench on ext4 on UMA showed
roughly 10% reduction in the number of CPU migrations and it is a case
where interrupts are frequent for IO competions. In most cases, the
difference in performance is quite small but variability is often
reduced. For example, this is the result for pgbench running on a UMA
machine with different numbers of clients.

                          4.15.0-rc9             4.15.0-rc9
                            baseline              waprev-v1
 Hmean     1     22096.28 (   0.00%)    22734.86 (   2.89%)
 Hmean     4     74633.42 (   0.00%)    75496.77 (   1.16%)
 Hmean     7    115017.50 (   0.00%)   113030.81 (  -1.73%)
 Hmean     12   126209.63 (   0.00%)   126613.40 (   0.32%)
 Hmean     16   131886.91 (   0.00%)   130844.35 (  -0.79%)
 Stddev    1       636.38 (   0.00%)      417.11 (  34.46%)
 Stddev    4       614.64 (   0.00%)      583.24 (   5.11%)
 Stddev    7       542.46 (   0.00%)      435.45 (  19.73%)
 Stddev    12      173.93 (   0.00%)      171.50 (   1.40%)
 Stddev    16      671.42 (   0.00%)      680.30 (  -1.32%)
 CoeffVar  1         2.88 (   0.00%)        1.83 (  36.26%)

Note that the different in performance is marginal but for low utilisation,
there is less variability.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180130104555.4125-4-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      806486c3
    • M
      sched/fair: Restructure wake_affine*() to return a CPU id · 3b76c4a3
      由 Mel Gorman 提交于 
      This is a preparation patch that has wake_affine*() return a CPU ID instead of
a boolean. The intent is to allow the wake_affine() helpers to be avoided
if a decision is already made. This patch has no functional change.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180130104555.4125-3-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      3b76c4a3
    • M
      sched/fair: Remove unnecessary parameters from wake_affine_idle() · 89a55f56
      由 Mel Gorman 提交于 
      wake_affine_idle() takes parameters it never uses so clean it up.
Signed-off-by: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180130104555.4125-2-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>
      89a55f56