Skip to content

cloud-kernel
cloud-kernel

openanolis / cloud-kernel
1 年多 前同步成功

通知 160
Star 36
Fork 7
cloud-kernel
cloud-kernel
切换分支/标签
  1. 11 6月, 2019 2 次提交
    • J
      mm: workingset: tell cache transitions from workingset thrashing · 20403665
      由 Johannes Weiner 提交于 
      commit 1899ad18c6072d689896badafb81267b0a1092a4 upstream.

Refaults happen during transitions between workingsets as well as in-place
thrashing.  Knowing the difference between the two has a range of
applications, including measuring the impact of memory shortage on the
system performance, as well as the ability to smarter balance pressure
between the filesystem cache and the swap-backed workingset.

During workingset transitions, inactive cache refaults and pushes out
established active cache.  When that active cache isn't stale, however,
and also ends up refaulting, that's bonafide thrashing.

Introduce a new page flag that tells on eviction whether the page has been
active or not in its lifetime.  This bit is then stored in the shadow
entry, to classify refaults as transitioning or thrashing.

How many page->flags does this leave us with on 32-bit?

	20 bits are always page flags

	21 if you have an MMU

	23 with the zone bits for DMA, Normal, HighMem, Movable

	29 with the sparsemem section bits

	30 if PAE is enabled

	31 with this patch.

So on 32-bit PAE, that leaves 1 bit for distinguishing two NUMA nodes.  If
that's not enough, the system can switch to discontigmem and re-gain the 6
or 7 sparsemem section bits.

Link: http://lkml.kernel.org/r/20180828172258.3185-3-hannes@cmpxchg.orgSigned-off-by: NJohannes Weiner <hannes@cmpxchg.org>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: NDaniel Drake <drake@endlessm.com>
Tested-by: NSuren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: NJoseph Qi <joseph.qi@linux.alibaba.com>
Acked-by: NCaspar Zhang <caspar@linux.alibaba.com>
      20403665
    • J
      mm: workingset: don't drop refault information prematurely · 63a30543
      由 Johannes Weiner 提交于 
      commit 95f9ab2d596e8cbb388315e78c82b9a131bf2928 upstream.

Patch series "psi: pressure stall information for CPU, memory, and IO", v4.

		Overview

PSI reports the overall wallclock time in which the tasks in a system (or
cgroup) wait for (contended) hardware resources.

This helps users understand the resource pressure their workloads are
under, which allows them to rootcause and fix throughput and latency
problems caused by overcommitting, underprovisioning, suboptimal job
placement in a grid; as well as anticipate major disruptions like OOM.

		Real-world applications

We're using the data collected by PSI (and its previous incarnation,
memdelay) quite extensively at Facebook, and with several success stories.

One usecase is avoiding OOM hangs/livelocks.  The reason these happen is
because the OOM killer is triggered by reclaim not being able to free
pages, but with fast flash devices there is *always* some clean and
uptodate cache to reclaim; the OOM killer never kicks in, even as tasks
spend 90% of the time thrashing the cache pages of their own executables.
There is no situation where this ever makes sense in practice.  We wrote a
<100 line POC python script to monitor memory pressure and kill stuff way
before such pathological thrashing leads to full system losses that would
require forcible hard resets.

We've since extended and deployed this code into other places to guarantee
latency and throughput SLAs, since they're usually violated way before the
kernel OOM killer would ever kick in.

It is available here: https://github.com/facebookincubator/oomd

Eventually we probably want to trigger the in-kernel OOM killer based on
extreme sustained pressure as well, so that Linux can avoid memory
livelocks - which technically aren't deadlocks, but to the user
indistinguishable from them - out of the box.  We'd continue using OOMD as
the first line of defense to ensure workload health and implement complex
kill policies that are beyond the scope of the kernel.

We also use PSI memory pressure for loadshedding.  Our batch job
infrastructure used to use heuristics based on various VM stats to
anticipate OOM situations, with lackluster success.  We switched it to PSI
and managed to anticipate and avoid OOM kills and lockups fairly reliably.
The reduction of OOM outages in the worker pool raised the pool's
aggregate productivity, and we were able to switch that service to smaller
machines.

Lastly, we use cgroups to isolate a machine's main workload from
maintenance crap like package upgrades, logging, configuration, as well as
to prevent multiple workloads on a machine from stepping on each others'
toes.  We were not able to configure this properly without the pressure
metrics; we would see latency or bandwidth drops, but it would often be
hard to impossible to rootcause it post-mortem.

We now log and graph pressure for the containers in our fleet and can
trivially link latency spikes and throughput drops to shortages of
specific resources after the fact, and fix the job config/scheduling.

PSI has also received testing, feedback, and feature requests from Android
and EndlessOS for the purpose of low-latency OOM killing, to intervene in
pressure situations before the UI starts hanging.

		How do you use this feature?

A kernel with CONFIG_PSI=y will create a /proc/pressure directory with 3
files: cpu, memory, and io.  If using cgroup2, cgroups will also have
cpu.pressure, memory.pressure and io.pressure files, which simply
aggregate task stalls at the cgroup level instead of system-wide.

The cpu file contains one line:

	some avg10=2.04 avg60=0.75 avg300=0.40 total=157656722

The averages give the percentage of walltime in which one or more tasks
are delayed on the runqueue while another task has the CPU.  They're
recent averages over 10s, 1m, 5m windows, so you can tell short term
trends from long term ones, similarly to the load average.

The total= value gives the absolute stall time in microseconds.  This
allows detecting latency spikes that might be too short to sway the
running averages.  It also allows custom time averaging in case the
10s/1m/5m windows aren't adequate for the usecase (or are too coarse with
future hardware).

What to make of this "some" metric?  If CPU utilization is at 100% and CPU
pressure is 0, it means the system is perfectly utilized, with one
runnable thread per CPU and nobody waiting.  At two or more runnable tasks
per CPU, the system is 100% overcommitted and the pressure average will
indicate as much.  From a utilization perspective this is a great state of
course: no CPU cycles are being wasted, even when 50% of the threads were
to go idle (as most workloads do vary).  From the perspective of the
individual job it's not great, however, and they would do better with more
resources.  Depending on what your priority and options are, raised "some"
numbers may or may not require action.

The memory file contains two lines:

some avg10=70.24 avg60=68.52 avg300=69.91 total=3559632828
full avg10=57.59 avg60=58.06 avg300=60.38 total=3300487258

The some line is the same as for cpu, the time in which at least one task
is stalled on the resource.  In the case of memory, this includes waiting
on swap-in, page cache refaults and page reclaim.

The full line, however, indicates time in which *nobody* is using the CPU
productively due to pressure: all non-idle tasks are waiting for memory in
one form or another.  Significant time spent in there is a good trigger
for killing things, moving jobs to other machines, or dropping incoming
requests, since neither the jobs nor the machine overall are making too
much headway.

The io file is similar to memory.  Because the block layer doesn't have a
concept of hardware contention right now (how much longer is my IO request
taking due to other tasks?), it reports CPU potential lost on all IO
delays, not just the potential lost due to competition.

		FAQ

Q: How is PSI's CPU component different from the load average?

A: There are several quirks in the load average that make it hard to
   impossible to tell how overcommitted the CPU really is.

   1. The load average is reported as a raw number of active tasks.
      You need to know how many CPUs there are in the system, how many
      CPUs the workload is allowed to use, then think about what the
      proportion between load and the number of CPUs mean for the
      tasks trying to run.

      PSI reports the percentage of wallclock time in which tasks are
      waiting for a CPU to run on. It doesn't matter how many CPUs are
      present or usable. The number always tells the quality of life
      of tasks in the system or in a particular cgroup.

   2. The shortest averaging window is 1m, which is extremely coarse,
      and it's sampled in 5s intervals. A *lot* can happen on a CPU in
      5 seconds. This *may* be able to identify persistent long-term
      trends and very clear and obvious overloads, but it's unusable
      for latency spikes and more subtle overutilization.

      PSI's shortest window is 10s. It also exports the cumulative
      stall times (in microseconds) of synchronously recorded events.

   3. On Linux, the load average for historical reasons includes all
      TASK_UNINTERRUPTIBLE tasks. This gives a broader sense of how
      busy the system is, but on the flipside it doesn't distinguish
      whether tasks are likely to contend over the CPU or IO - which
      obviously requires very different interventions from a sys admin
      or a job scheduler.

      PSI reports independent metrics for CPU and IO. You can tell
      which resource is making the tasks wait, but in conjunction
      still see how overloaded the system is overall.

Q: What's the cost / performance impact of this feature?

A: PSI's primary cost is in the scheduler, in particular task wakeups
   and sleeps.

   I benchmarked this code using Facebook's two most scheduling
   sensitive workloads: memcache and webserver. They handle a ton of
   small requests - lots of wakeups and sleeps with little actual work
   in between - so they tend to be canaries for scheduler regressions.

   In the tests, the boxes were handling live traffic over the course
   of several hours. Half the machines, the control, ran with
   CONFIG_PSI=n.

   For memcache I used eight machines total. They're 2-socket, 14
   core, 56 thread boxes. The test runs for half the test period,
   flips the test and control kernels on the hardware to rule out HW
   factors, DC location etc., then runs the other half of the test.

   For the webservers, I used 32 machines total. They're single
   socket, 16 core, 32 thread machines.

   During the memcache test, CPU load was nopsi=78.05% psi=78.98% in
   the first half and nopsi=77.52% psi=78.25%, so PSI added between
   0.7 and 0.9 percentage points to the CPU load, a difference of
   about 1%.

   UPDATE: I re-ran this test with the v3 version of this patch set
   and the CPU utilization was equivalent between test and control.

   UPDATE: v4 is on par with v3.

   As far as end-to-end request latency from the client perspective
   goes, we don't sample those finely enough to capture the requests
   going to those particular machines during the test, but we know the
   p50 turnaround time in this workload is 54us, and perf bench sched
   pipe on those machines show nopsi=5.232666 us/op and psi=5.587347
   us/op, so this doesn't add much here either.

   The profile for the pipe benchmark shows:

        0.87%  sched-pipe  [kernel.vmlinux]    [k] psi_group_change
        0.83%  perf.real   [kernel.vmlinux]    [k] psi_group_change
        0.82%  perf.real   [kernel.vmlinux]    [k] psi_task_change
        0.58%  sched-pipe  [kernel.vmlinux]    [k] psi_task_change

   The webserver load is running inside 4 nested cgroup levels. The
   CPU load with both nopsi and psi kernels was indistinguishable at
   81%.

   For comparison, we had to disable the cgroup cpu controller on the
   webservers because it added 4 percentage points to the CPU% during
   this same exact test.

   Versions of this accounting code now run on 80% of our fleet. None
   of our workloads have reported regressions during the rollout.

Daniel Drake said:

: I just retested the latest version at
: http://git.cmpxchg.org/cgit.cgi/linux-psi.git (Linux 4.18) and the results
: are great.
:
: Test setup:
: Endless OS
: GeminiLake N4200 low end laptop
: 2GB RAM
: swap (and zram swap) disabled
:
: Baseline test: open a handful of large-ish apps and several website
: tabs in Google Chrome.
:
: Results: after a couple of minutes, system is excessively thrashing, mouse
: cursor can barely be moved, UI is not responding to mouse clicks, so it's
: impractical to recover from this situation as an ordinary user
:
: Add my simple killer:
: https://gist.github.com/dsd/a8988bf0b81a6163475988120fe8d9cd
:
: Results: when the thrashing causes the UI to become sluggish, the killer
: steps in and kills something (usually a chrome tab), and the system
: remains usable.  I repeatedly opened more apps and more websites over a 15
: minute period but I wasn't able to get the system to a point of UI
: unresponsiveness.

Suren said:

: Backported to 4.9 and retested on ARMv8 8 code system running Android.
: Signals behave as expected reacting to memory pressure, no jumps in
: "total" counters that would indicate an overflow/underflow issues.  Nicely
: done!

This patch (of 9):

If we keep just enough refault information to match the *current* page
cache during reclaim time, we could lose a lot of events when there is
only a temporary spike in non-cache memory consumption that pushes out all
the cache.  Once cache comes back, we won't see those refaults.  They
might not be actionable for LRU aging, but we want to know about them for
measuring memory pressure.

[hannes@cmpxchg.org: switch to NUMA-aware lru and slab counters]
  Link: http://lkml.kernel.org/r/20181009184732.762-2-hannes@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-2-hannes@cmpxchg.orgSigned-off-by: NJohannes Weiner <jweiner@fb.com>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: NRik van Riel <riel@surriel.com>
Tested-by: NDaniel Drake <drake@endlessm.com>
Tested-by: NSuren Baghdasaryan <surenb@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Christopher Lameter <cl@linux.com>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: NJoseph Qi <joseph.qi@linux.alibaba.com>
Acked-by: NCaspar Zhang <caspar@linux.alibaba.com>
      63a30543
  3. 10 6月, 2019 3 次提交
  5. 04 6月, 2019 35 次提交