| Age | Commit message (Collapse) | Author |
|---|
|
Tracing the runtime delta makes sense, observer can sum over time.
Tracing the absolute vruntime makes less sense, inconsistent:
absolute-vs-delta, but also vruntime delta can be computed from
runtime delta.
Removing the vruntime thing also makes the two tracepoint sites
identical, allowing to unify the code in a later patch.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
|
|
All classes use sched_entity::exec_start to track runtime and have
copies of the exact same code around to compute runtime.
Collapse all that.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Daniel Bristot de Oliveira <bristot@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Phil Auld <pauld@redhat.com>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Link: https://lkml.kernel.org/r/54d148a144f26d9559698c4dd82d8859038a7380.1699095159.git.bristot@kernel.org
|
|
Since the RB-tree is now sorted by deadline, let's first try the
leftmost entity which has the earliest virtual deadline. I've done
some benchmarks to see its effectiveness.
All the benchmarks are done inside a normal cpu cgroup in a clean
environment with cpu turbo disabled, on a dual-CPU Intel Xeon(R)
Platinum 8260 with 2 NUMA nodes each of which has 24C/48T.
hackbench: process/thread + pipe/socket + 1/2/4/8 groups
netperf: TCP/UDP + STREAM/RR + 24/48/72/96/192 threads
tbench: loopback 24/48/72/96/192 threads
schbench: 1/2/4/8 mthreads
direct: cfs_rq has only one entity
parity: RUN_TO_PARITY
fast: O(1) fastpath
slow: heap search
(%) direct parity fast slow
hackbench 92.95 2.02 4.91 0.12
netperf 68.08 6.60 24.18 1.14
tbench 67.55 11.22 20.61 0.62
schbench 69.91 2.65 25.73 1.71
The above results indicate that this fastpath really makes task
selection more efficient.
Signed-off-by: Abel Wu <wuyun.abel@bytedance.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20231115033647.80785-4-wuyun.abel@bytedance.com
|
|
Sort the task timeline by virtual deadline and keep the min_vruntime
in the augmented tree, so we can avoid doubling the worst case cost
and make full use of the cached leftmost node to enable O(1) fastpath
picking in next patch.
Signed-off-by: Abel Wu <wuyun.abel@bytedance.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20231115033647.80785-3-wuyun.abel@bytedance.com
|
|
Since commit fc137c0ddab2 ("sched/numa: enhance vma scanning logic")
NUMA Balancing allows updating PTEs to trap NUMA hinting faults if the
task had previously accessed VMA. However unconditional scan of VMAs are
allowed during initial phase of VMA creation until process's
mm numa_scan_seq reaches 2 even though current task had not accessed VMA.
Rationale:
- Without initial scan subsequent PTE update may never happen.
- Give fair opportunity to all the VMAs to be scanned and subsequently
understand the access pattern of all the VMAs.
But it has a corner case where, if a VMA is created after some time,
process's mm numa_scan_seq could be already greater than 2.
For e.g., values of mm numa_scan_seq when VMAs are created by running
mmtest autonuma benchmark briefly looks like:
start_seq=0 : 459
start_seq=2 : 138
start_seq=3 : 144
start_seq=4 : 8
start_seq=8 : 1
start_seq=9 : 1
This results in no unconditional PTE updates for those VMAs created after
some time.
Fix:
- Note down the initial value of mm numa_scan_seq in per VMA start_seq.
- Allow unconditional scan till start_seq + 2.
Result:
SUT: AMD EPYC Milan with 2 NUMA nodes 256 cpus.
base kernel: upstream 6.6-rc6 with Mels patches [1] applied.
kernbench
========== base patched %gain
Amean elsp-128 165.09 ( 0.00%) 164.78 * 0.19%*
Duration User 41404.28 41375.08
Duration System 9862.22 9768.48
Duration Elapsed 519.87 518.72
Ops NUMA PTE updates 1041416.00 831536.00
Ops NUMA hint faults 263296.00 220966.00
Ops NUMA pages migrated 258021.00 212769.00
Ops AutoNUMA cost 1328.67 1114.69
autonumabench
NUMA01_THREADLOCAL
==================
Amean elsp-NUMA01_THREADLOCAL 81.79 (0.00%) 67.74 * 17.18%*
Duration User 54832.73 47379.67
Duration System 75.00 185.75
Duration Elapsed 576.72 476.09
Ops NUMA PTE updates 394429.00 11121044.00
Ops NUMA hint faults 1001.00 8906404.00
Ops NUMA pages migrated 288.00 2998694.00
Ops AutoNUMA cost 7.77 44666.84
Signed-off-by: Raghavendra K T <raghavendra.kt@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Link: https://lore.kernel.org/r/2ea7cbce80ac7c62e90cbfb9653a7972f902439f.1697816692.git.raghavendra.kt@amd.com
|
|
should_we_balance is called for the decision to do load-balancing.
When sched ticks invoke this function, only one CPU should return
true. However, in the current code, two CPUs can return true. The
following situation, where b means busy and i means idle, is an
example, because CPU 0 and CPU 2 return true.
[0, 1] [2, 3]
b b i b
This fix checks if there exists an idle CPU with busy sibling(s)
after looking for a CPU on an idle core. If some idle CPUs with busy
siblings are found, just the first one should do load-balancing.
Fixes: b1bfeab9b002 ("sched/fair: Consider the idle state of the whole core for load balance")
Signed-off-by: Keisuke Nishimura <keisuke.nishimura@inria.fr>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Chen Yu <yu.c.chen@intel.com>
Reviewed-by: Shrikanth Hegde <sshegde@linux.vnet.ibm.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20231031133821.1570861-1-keisuke.nishimura@inria.fr
|
|
vruntime of the (on_rq && !0-lag) entity needs to be adjusted when
it gets re-weighted, and the calculations can be simplified based
on the fact that re-weight won't change the w-average of all the
entities. Please check the proofs in comments.
But adjusting vruntime can also cause position change in RB-tree
hence require re-queue to fix up which might be costly. This might
be avoided by deferring adjustment to the time the entity actually
leaves tree (dequeue/pick), but that will negatively affect task
selection and probably not good enough either.
Fixes: 147f3efaa241 ("sched/fair: Implement an EEVDF-like scheduling policy")
Signed-off-by: Abel Wu <wuyun.abel@bytedance.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20231107090510.71322-2-wuyun.abel@bytedance.com
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm
Pull MM updates from Andrew Morton:
"Many singleton patches against the MM code. The patch series which are
included in this merge do the following:
- Kemeng Shi has contributed some compation maintenance work in the
series 'Fixes and cleanups to compaction'
- Joel Fernandes has a patchset ('Optimize mremap during mutual
alignment within PMD') which fixes an obscure issue with mremap()'s
pagetable handling during a subsequent exec(), based upon an
implementation which Linus suggested
- More DAMON/DAMOS maintenance and feature work from SeongJae Park i
the following patch series:
mm/damon: misc fixups for documents, comments and its tracepoint
mm/damon: add a tracepoint for damos apply target regions
mm/damon: provide pseudo-moving sum based access rate
mm/damon: implement DAMOS apply intervals
mm/damon/core-test: Fix memory leaks in core-test
mm/damon/sysfs-schemes: Do DAMOS tried regions update for only one apply interval
- In the series 'Do not try to access unaccepted memory' Adrian
Hunter provides some fixups for the recently-added 'unaccepted
memory' feature. To increase the feature's checking coverage. 'Plug
a few gaps where RAM is exposed without checking if it is
unaccepted memory'
- In the series 'cleanups for lockless slab shrink' Qi Zheng has done
some maintenance work which is preparation for the lockless slab
shrinking code
- Qi Zheng has redone the earlier (and reverted) attempt to make slab
shrinking lockless in the series 'use refcount+RCU method to
implement lockless slab shrink'
- David Hildenbrand contributes some maintenance work for the rmap
code in the series 'Anon rmap cleanups'
- Kefeng Wang does more folio conversions and some maintenance work
in the migration code. Series 'mm: migrate: more folio conversion
and unification'
- Matthew Wilcox has fixed an issue in the buffer_head code which was
causing long stalls under some heavy memory/IO loads. Some cleanups
were added on the way. Series 'Add and use bdev_getblk()'
- In the series 'Use nth_page() in place of direct struct page
manipulation' Zi Yan has fixed a potential issue with the direct
manipulation of hugetlb page frames
- In the series 'mm: hugetlb: Skip initialization of gigantic tail
struct pages if freed by HVO' has improved our handling of gigantic
pages in the hugetlb vmmemmep optimizaton code. This provides
significant boot time improvements when significant amounts of
gigantic pages are in use
- Matthew Wilcox has sent the series 'Small hugetlb cleanups' - code
rationalization and folio conversions in the hugetlb code
- Yin Fengwei has improved mlock()'s handling of large folios in the
series 'support large folio for mlock'
- In the series 'Expose swapcache stat for memcg v1' Liu Shixin has
added statistics for memcg v1 users which are available (and
useful) under memcg v2
- Florent Revest has enhanced the MDWE (Memory-Deny-Write-Executable)
prctl so that userspace may direct the kernel to not automatically
propagate the denial to child processes. The series is named 'MDWE
without inheritance'
- Kefeng Wang has provided the series 'mm: convert numa balancing
functions to use a folio' which does what it says
- In the series 'mm/ksm: add fork-exec support for prctl' Stefan
Roesch makes is possible for a process to propagate KSM treatment
across exec()
- Huang Ying has enhanced memory tiering's calculation of memory
distances. This is used to permit the dax/kmem driver to use 'high
bandwidth memory' in addition to Optane Data Center Persistent
Memory Modules (DCPMM). The series is named 'memory tiering:
calculate abstract distance based on ACPI HMAT'
- In the series 'Smart scanning mode for KSM' Stefan Roesch has
optimized KSM by teaching it to retain and use some historical
information from previous scans
- Yosry Ahmed has fixed some inconsistencies in memcg statistics in
the series 'mm: memcg: fix tracking of pending stats updates
values'
- In the series 'Implement IOCTL to get and optionally clear info
about PTEs' Peter Xu has added an ioctl to /proc/<pid>/pagemap
which permits us to atomically read-then-clear page softdirty
state. This is mainly used by CRIU
- Hugh Dickins contributed the series 'shmem,tmpfs: general
maintenance', a bunch of relatively minor maintenance tweaks to
this code
- Matthew Wilcox has increased the use of the VMA lock over
file-backed page faults in the series 'Handle more faults under the
VMA lock'. Some rationalizations of the fault path became possible
as a result
- In the series 'mm/rmap: convert page_move_anon_rmap() to
folio_move_anon_rmap()' David Hildenbrand has implemented some
cleanups and folio conversions
- In the series 'various improvements to the GUP interface' Lorenzo
Stoakes has simplified and improved the GUP interface with an eye
to providing groundwork for future improvements
- Andrey Konovalov has sent along the series 'kasan: assorted fixes
and improvements' which does those things
- Some page allocator maintenance work from Kemeng Shi in the series
'Two minor cleanups to break_down_buddy_pages'
- In thes series 'New selftest for mm' Breno Leitao has developed
another MM self test which tickles a race we had between madvise()
and page faults
- In the series 'Add folio_end_read' Matthew Wilcox provides cleanups
and an optimization to the core pagecache code
- Nhat Pham has added memcg accounting for hugetlb memory in the
series 'hugetlb memcg accounting'
- Cleanups and rationalizations to the pagemap code from Lorenzo
Stoakes, in the series 'Abstract vma_merge() and split_vma()'
- Audra Mitchell has fixed issues in the procfs page_owner code's new
timestamping feature which was causing some misbehaviours. In the
series 'Fix page_owner's use of free timestamps'
- Lorenzo Stoakes has fixed the handling of new mappings of sealed
files in the series 'permit write-sealed memfd read-only shared
mappings'
- Mike Kravetz has optimized the hugetlb vmemmap optimization in the
series 'Batch hugetlb vmemmap modification operations'
- Some buffer_head folio conversions and cleanups from Matthew Wilcox
in the series 'Finish the create_empty_buffers() transition'
- As a page allocator performance optimization Huang Ying has added
automatic tuning to the allocator's per-cpu-pages feature, in the
series 'mm: PCP high auto-tuning'
- Roman Gushchin has contributed the patchset 'mm: improve
performance of accounted kernel memory allocations' which improves
their performance by ~30% as measured by a micro-benchmark
- folio conversions from Kefeng Wang in the series 'mm: convert page
cpupid functions to folios'
- Some kmemleak fixups in Liu Shixin's series 'Some bugfix about
kmemleak'
- Qi Zheng has improved our handling of memoryless nodes by keeping
them off the allocation fallback list. This is done in the series
'handle memoryless nodes more appropriately'
- khugepaged conversions from Vishal Moola in the series 'Some
khugepaged folio conversions'"
[ bcachefs conflicts with the dynamically allocated shrinkers have been
resolved as per Stephen Rothwell in
https://lore.kernel.org/all/20230913093553.4290421e@canb.auug.org.au/
with help from Qi Zheng.
The clone3 test filtering conflict was half-arsed by yours truly ]
* tag 'mm-stable-2023-11-01-14-33' of git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm: (406 commits)
mm/damon/sysfs: update monitoring target regions for online input commit
mm/damon/sysfs: remove requested targets when online-commit inputs
selftests: add a sanity check for zswap
Documentation: maple_tree: fix word spelling error
mm/vmalloc: fix the unchecked dereference warning in vread_iter()
zswap: export compression failure stats
Documentation: ubsan: drop "the" from article title
mempolicy: migration attempt to match interleave nodes
mempolicy: mmap_lock is not needed while migrating folios
mempolicy: alloc_pages_mpol() for NUMA policy without vma
mm: add page_rmappable_folio() wrapper
mempolicy: remove confusing MPOL_MF_LAZY dead code
mempolicy: mpol_shared_policy_init() without pseudo-vma
mempolicy trivia: use pgoff_t in shared mempolicy tree
mempolicy trivia: slightly more consistent naming
mempolicy trivia: delete those ancient pr_debug()s
mempolicy: fix migrate_pages(2) syscall return nr_failed
kernfs: drop shared NUMA mempolicy hooks
hugetlbfs: drop shared NUMA mempolicy pretence
mm/damon/sysfs-test: add a unit test for damon_sysfs_set_targets()
...
|
|
Convert to use folio_xchg_last_cpupid() in should_numa_migrate_memory().
Link: https://lkml.kernel.org/r/20231018140806.2783514-14-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
|
|
Convert to use folio_xchg_access_time() in numa_hint_fault_latency().
Link: https://lkml.kernel.org/r/20231018140806.2783514-9-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
|
|
SIS_UTIL seems to work well, lets remove the old thing.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20231020134337.GD33965@noisy.programming.kicks-ass.net
|
|
cluster wakeup
Chen Yu reports a hackbench regression of cluster wakeup when
hackbench threads equal to the CPU number [1]. Analysis shows
it's because we wake up more on the target CPU even if the
prev_cpu is a good wakeup candidate and leads to the decrease
of the CPU utilization.
Generally if the task's prev_cpu is idle we'll wake up the task
on it without scanning. On cluster machines we'll try to wake up
the task in the same cluster of the target for better cache
affinity, so if the prev_cpu is idle but not sharing the same
cluster with the target we'll still try to find an idle CPU within
the cluster. This will improve the performance at low loads on
cluster machines. But in the issue above, if the prev_cpu is idle
but not in the cluster with the target CPU, we'll try to scan an
idle one in the cluster. But since the system is busy, we're
likely to fail the scanning and use target instead, even if
the prev_cpu is idle. Then leads to the regression.
This patch solves this in 2 steps:
o record the prev_cpu/recent_used_cpu if they're good wakeup
candidates but not sharing the cluster with the target.
o on scanning failure use the prev_cpu/recent_used_cpu if
they're recorded as idle
[1] https://lore.kernel.org/all/ZGzDLuVaHR1PAYDt@chenyu5-mobl1/
Closes: https://lore.kernel.org/all/ZGsLy83wPIpamy6x@chenyu5-mobl1/
Reported-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Tested-and-reviewed-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20231019033323.54147-4-yangyicong@huawei.com
|
|
For platforms having clusters like Kunpeng920, CPUs within the same cluster
have lower latency when synchronizing and accessing shared resources like
cache. Thus, this patch tries to find an idle cpu within the cluster of the
target CPU before scanning the whole LLC to gain lower latency. This
will be implemented in 2 steps in select_idle_sibling():
1. When the prev_cpu/recent_used_cpu are good wakeup candidates, use them
if they're sharing cluster with the target CPU. Otherwise trying to
scan for an idle CPU in the target's cluster.
2. Scanning the cluster prior to the LLC of the target CPU for an
idle CPU to wakeup.
Testing has been done on Kunpeng920 by pinning tasks to one numa and two
numa. On Kunpeng920, Each numa has 8 clusters and each cluster has 4 CPUs.
With this patch, We noticed enhancement on tbench and netperf within one
numa or cross two numa on top of tip-sched-core commit
9b46f1abc6d4 ("sched/debug: Print 'tgid' in sched_show_task()")
tbench results (node 0):
baseline patched
1: 327.2833 372.4623 ( 13.80%)
4: 1320.5933 1479.8833 ( 12.06%)
8: 2638.4867 2921.5267 ( 10.73%)
16: 5282.7133 5891.5633 ( 11.53%)
32: 9810.6733 9877.3400 ( 0.68%)
64: 7408.9367 7447.9900 ( 0.53%)
128: 6203.2600 6191.6500 ( -0.19%)
tbench results (node 0-1):
baseline patched
1: 332.0433 372.7223 ( 12.25%)
4: 1325.4667 1477.6733 ( 11.48%)
8: 2622.9433 2897.9967 ( 10.49%)
16: 5218.6100 5878.2967 ( 12.64%)
32: 10211.7000 11494.4000 ( 12.56%)
64: 13313.7333 16740.0333 ( 25.74%)
128: 13959.1000 14533.9000 ( 4.12%)
netperf results TCP_RR (node 0):
baseline patched
1: 76546.5033 90649.9867 ( 18.42%)
4: 77292.4450 90932.7175 ( 17.65%)
8: 77367.7254 90882.3467 ( 17.47%)
16: 78519.9048 90938.8344 ( 15.82%)
32: 72169.5035 72851.6730 ( 0.95%)
64: 25911.2457 25882.2315 ( -0.11%)
128: 10752.6572 10768.6038 ( 0.15%)
netperf results TCP_RR (node 0-1):
baseline patched
1: 76857.6667 90892.2767 ( 18.26%)
4: 78236.6475 90767.3017 ( 16.02%)
8: 77929.6096 90684.1633 ( 16.37%)
16: 77438.5873 90502.5787 ( 16.87%)
32: 74205.6635 88301.5612 ( 19.00%)
64: 69827.8535 71787.6706 ( 2.81%)
128: 25281.4366 25771.3023 ( 1.94%)
netperf results UDP_RR (node 0):
baseline patched
1: 96869.8400 110800.8467 ( 14.38%)
4: 97744.9750 109680.5425 ( 12.21%)
8: 98783.9863 110409.9637 ( 11.77%)
16: 99575.0235 110636.2435 ( 11.11%)
32: 95044.7250 97622.8887 ( 2.71%)
64: 32925.2146 32644.4991 ( -0.85%)
128: 12859.2343 12824.0051 ( -0.27%)
netperf results UDP_RR (node 0-1):
baseline patched
1: 97202.4733 110190.1200 ( 13.36%)
4: 95954.0558 106245.7258 ( 10.73%)
8: 96277.1958 105206.5304 ( 9.27%)
16: 97692.7810 107927.2125 ( 10.48%)
32: 79999.6702 103550.2999 ( 29.44%)
64: 80592.7413 87284.0856 ( 8.30%)
128: 27701.5770 29914.5820 ( 7.99%)
Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so the SMT branch
in the code has not been tested but it supposed to work.
Chen Yu also noticed this will improve the performance of tbench and
netperf on a 24 CPUs Jacobsville machine, there are 4 CPUs in one
cluster sharing L2 Cache.
[https://lore.kernel.org/lkml/Ytfjs+m1kUs0ScSn@worktop.programming.kicks-ass.net]
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Barry Song <song.bao.hua@hisilicon.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>
Reviewed-by: Chen Yu <yu.c.chen@intel.com>
Reviewed-by: Gautham R. Shenoy <gautham.shenoy@amd.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-and-reviewed-by: Chen Yu <yu.c.chen@intel.com>
Tested-by: Yicong Yang <yangyicong@hisilicon.com>
Link: https://lkml.kernel.org/r/20231019033323.54147-3-yangyicong@huawei.com
|
|
Pick up recent sched/urgent fixes merged upstream.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
|
|
The 'curr' argument of pick_next_entity() has become unused after
the EEVDF changes.
[ mingo: Updated the changelog. ]
Signed-off-by: Yiwei Lin <s921975628@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20231020055617.42064-1-s921975628@gmail.com
|
|
How ILB is triggered without IPIs is cryptic. Out of mercy for future
code readers, document it in code comments.
The comments are derived from a discussion with Vincent in a past
review.
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20231020014031.919742-2-joel@joelfernandes.org
|
|
./kernel/sched/fair.c: linux/sched/cond_resched.h is included more than once.
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Jiapeng Chong <jiapeng.chong@linux.alibaba.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20231018062759.44375-1-jiapeng.chong@linux.alibaba.com
Closes: https://bugzilla.openanolis.cn/show_bug.cgi?id=6907
|
|
Because someone is a flaming idiot... and forgot we have current as
se->on_rq but not actually in the tree itself, and walking rb_parent()
on an entry not in the tree is 'funky' and KASAN complains.
Fixes: 8dafa9d0eb1a ("sched/eevdf: Fix min_deadline heap integrity")
Reported-by: 0599jiangyc@gmail.com
Reported-by: Dmitry Safonov <0x7f454c46@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Dmitry Safonov <0x7f454c46@gmail.com>
Link: https://bugzilla.kernel.org/show_bug.cgi?id=218020
Link: https://lkml.kernel.org/r/CAJwJo6ZGXO07%3DQvW4fgQfbsDzQPs9xj5sAQ1zp%3DmAyPMNbHYww%40mail.gmail.com
|
|
The cpupid (or access time) is stored in the head page for THP, so it is
safely to make should_numa_migrate_memory() and numa_hint_fault_latency()
to take a folio. This is in preparation for large folio numa balancing.
Link: https://lkml.kernel.org/r/20230921074417.24004-7-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
|
|
Kuyo reported sporadic failures on a sched_setaffinity() vs CPU
hotplug stress-test -- notably affine_move_task() remains stuck in
wait_for_completion(), leading to a hung-task detector warning.
Specifically, it was reported that stop_one_cpu_nowait(.fn =
migration_cpu_stop) returns false -- this stopper is responsible for
the matching complete().
The race scenario is:
CPU0 CPU1
// doing _cpu_down()
__set_cpus_allowed_ptr()
task_rq_lock();
takedown_cpu()
stop_machine_cpuslocked(take_cpu_down..)
<PREEMPT: cpu_stopper_thread()
MULTI_STOP_PREPARE
...
__set_cpus_allowed_ptr_locked()
affine_move_task()
task_rq_unlock();
<PREEMPT: cpu_stopper_thread()\>
ack_state()
MULTI_STOP_RUN
take_cpu_down()
__cpu_disable();
stop_machine_park();
stopper->enabled = false;
/>
/>
stop_one_cpu_nowait(.fn = migration_cpu_stop);
if (stopper->enabled) // false!!!
That is, by doing stop_one_cpu_nowait() after dropping rq-lock, the
stopper thread gets a chance to preempt and allows the cpu-down for
the target CPU to complete.
OTOH, since stop_one_cpu_nowait() / cpu_stop_queue_work() needs to
issue a wakeup, it must not be ran under the scheduler locks.
Solve this apparent contradiction by keeping preemption disabled over
the unlock + queue_stopper combination:
preempt_disable();
task_rq_unlock(...);
if (!stop_pending)
stop_one_cpu_nowait(...)
preempt_enable();
This respects the lock ordering contraints while still avoiding the
above race. That is, if we find the CPU is online under rq-lock, the
targeted stop_one_cpu_nowait() must succeed.
Apply this pattern to all similar stop_one_cpu_nowait() invocations.
Fixes: 6d337eab041d ("sched: Fix migrate_disable() vs set_cpus_allowed_ptr()")
Reported-by: "Kuyo Chang (張建文)" <Kuyo.Chang@mediatek.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: "Kuyo Chang (張建文)" <Kuyo.Chang@mediatek.com>
Link: https://lkml.kernel.org/r/20231010200442.GA16515@noisy.programming.kicks-ass.net
|
|
While reworking the x86 topology code Thomas tripped over creating a 'DIE' domain
for the package mask. :-)
Since these names are CONFIG_SCHED_DEBUG=y only, rename them to make the
name less ambiguous.
[ Shrikanth Hegde: rename on s390 as well. ]
[ Valentin Schneider: also rename it in the comments. ]
[ mingo: port to recent kernels & find all remaining occurances. ]
Reported-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Valentin Schneider <vschneid@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Heiko Carstens <hca@linux.ibm.com>
Acked-by: Gautham R. Shenoy <gautham.shenoy@amd.com>
Acked-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20230712141056.GI3100107@hirez.programming.kicks-ass.net
|
|
VMAs are skipped if there is no recent fault activity but this represents
a chicken-and-egg problem as there may be no fault activity if the PTEs
are never updated to trap NUMA hints. There is an indirect reliance on
scanning to be forced early in the lifetime of a task but this may fail
to detect changes in phase behaviour. Force inactive VMAs to be scanned
when all other eligible VMAs have been updated within the same scan
sequence.
Test results in general look good with some changes in performance, both
negative and positive, depending on whether the additional scanning and
faulting was beneficial or not to the workload. The autonuma benchmark
workload NUMA01_THREADLOCAL was picked for closer examination. The workload
creates two processes with numerous threads and thread-local storage that
is zero-filled in a loop. It exercises the corner case where unrelated
threads may skip VMAs that are thread-local to another thread and still
has some VMAs that inactive while the workload executes.
The VMA skipping activity frequency with and without the patch:
6.6.0-rc2-sched-numabtrace-v1
=============================
649 reason=scan_delay
9,094 reason=unsuitable
48,915 reason=shared_ro
143,919 reason=inaccessible
193,050 reason=pid_inactive
6.6.0-rc2-sched-numabselective-v1
=============================
146 reason=seq_completed
622 reason=ignore_pid_inactive
624 reason=scan_delay
6,570 reason=unsuitable
16,101 reason=shared_ro
27,608 reason=inaccessible
41,939 reason=pid_inactive
Note that with the patch applied, the PID activity is ignored
(ignore_pid_inactive) to ensure a VMA with some activity is completely
scanned. In addition, a small number of VMAs are scanned when no other
eligible VMA is available during a single scan window (seq_completed).
The number of times a VMA is skipped due to no PID activity from the
scanning task (pid_inactive) drops dramatically. It is expected that
this will increase the number of PTEs updated for NUMA hinting faults
as well as hinting faults but these represent PTEs that would otherwise
have been missed. The tradeoff is scan+fault overhead versus improving
locality due to migration.
On a 2-socket Cascade Lake test machine, the time to complete the
workload is as follows;
6.6.0-rc2 6.6.0-rc2
sched-numabtrace-v1 sched-numabselective-v1
Min elsp-NUMA01_THREADLOCAL 174.22 ( 0.00%) 117.64 ( 32.48%)
Amean elsp-NUMA01_THREADLOCAL 175.68 ( 0.00%) 123.34 * 29.79%*
Stddev elsp-NUMA01_THREADLOCAL 1.20 ( 0.00%) 4.06 (-238.20%)
CoeffVar elsp-NUMA01_THREADLOCAL 0.68 ( 0.00%) 3.29 (-381.70%)
Max elsp-NUMA01_THREADLOCAL 177.18 ( 0.00%) 128.03 ( 27.74%)
The time to complete the workload is reduced by almost 30%:
6.6.0-rc2 6.6.0-rc2
sched-numabtrace-v1 sched-numabselective-v1 /
Duration User 91201.80 63506.64
Duration System 2015.53 1819.78
Duration Elapsed 1234.77 868.37
In this specific case, system CPU time was not increased but it's not
universally true.
From vmstat, the NUMA scanning and fault activity is as follows;
6.6.0-rc2 6.6.0-rc2
sched-numabtrace-v1 sched-numabselective-v1
Ops NUMA base-page range updates 64272.00 26374386.00
Ops NUMA PTE updates 36624.00 55538.00
Ops NUMA PMD updates 54.00 51404.00
Ops NUMA hint faults 15504.00 75786.00
Ops NUMA hint local faults % 14860.00 56763.00
Ops NUMA hint local percent 95.85 74.90
Ops NUMA pages migrated 1629.00 6469222.00
Both the number of PTE updates and hint faults is dramatically
increased. While this is superficially unfortunate, it represents
ranges that were simply skipped without the patch. As a result
of the scanning and hinting faults, many more pages were also
migrated but as the time to completion is reduced, the overhead
is offset by the gain.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Raghavendra K T <raghavendra.kt@amd.com>
Link: https://lore.kernel.org/r/20231010083143.19593-7-mgorman@techsingularity.net
|
|
NUMA Balancing skips VMAs when the current task has not trapped a NUMA
fault within the VMA. If the VMA is skipped then mm->numa_scan_offset
advances and a task that is trapping faults within the VMA may never
fully update PTEs within the VMA.
Force tasks to update PTEs for partially scanned PTEs. The VMA will
be tagged for NUMA hints by some task but this removes some of the
benefit of tracking PID activity within a VMA. A follow-on patch
will mitigate this problem.
The test cases and machines evaluated did not trigger the corner case so
the performance results are neutral with only small changes within the
noise from normal test-to-test variance. However, the next patch makes
the corner case easier to trigger.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Raghavendra K T <raghavendra.kt@amd.com>
Link: https://lore.kernel.org/r/20231010083143.19593-6-mgorman@techsingularity.net
|
|
Recent NUMA hinting faulting activity is reset approximately every
VMA_PID_RESET_PERIOD milliseconds. However, if the current task has not
accessed a VMA then the reset check is missed and the reset is potentially
deferred forever. Check if the PID activity information should be reset
before checking if the current task recently trapped a NUMA hinting fault.
[ mgorman@techsingularity.net: Rewrite changelog ]
Suggested-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Raghavendra K T <raghavendra.kt@amd.com>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20231010083143.19593-5-mgorman@techsingularity.net
|
|
NUMA balancing skips or scans VMAs for a variety of reasons. In preparation
for completing scans of VMAs regardless of PID access, trace the reasons
why a VMA was skipped. In a later patch, the tracing will be used to track
if a VMA was forcibly scanned.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20231010083143.19593-4-mgorman@techsingularity.net
|
|
::next_pid_reset => ::pids_active_reset
The access_pids[] field name is somewhat ambiguous as no PIDs are accessed.
Similarly, it's not clear that next_pid_reset is related to access_pids[].
Rename the fields to more accurately reflect their purpose.
[ mingo: Rename in the comments too. ]
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20231010083143.19593-3-mgorman@techsingularity.net
|
|
Remove the rq::cpu_capacity_orig field and use arch_scale_cpu_capacity()
instead.
The scheduler uses 3 methods to get access to a CPU's max compute capacity:
- arch_scale_cpu_capacity(cpu) which is the default way to get a CPU's capacity.
- cpu_capacity_orig field which is periodically updated with
arch_scale_cpu_capacity().
- capacity_orig_of(cpu) which encapsulates rq->cpu_capacity_orig.
There is no real need to save the value returned by arch_scale_cpu_capacity()
in struct rq. arch_scale_cpu_capacity() returns:
- either a per_cpu variable.
- or a const value for systems which have only one capacity.
Remove rq::cpu_capacity_orig and use arch_scale_cpu_capacity() everywhere.
No functional changes.
Some performance tests on Arm64:
- small SMP device (hikey): no noticeable changes
- HMP device (RB5): hackbench shows minor improvement (1-2%)
- large smp (thx2): hackbench and tbench shows minor improvement (1%)
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Link: https://lore.kernel.org/r/20231009103621.374412-2-vincent.guittot@linaro.org
|
|
find_new_ilb()
find_new_ilb() returns nr_cpu_ids on failure - which is the usual
cpumask bitops return pattern, but is weird & unnecessary in this
context: not only is it a global variable, it it is a +1 out of
bounds CPU index and also has different signedness ...
Its only user, kick_ilb(), then checks the return against nr_cpu_ids
to decide to return. There's no other use.
So instead of this, use a standard -1 return on failure to find an
idle CPU, as the argument is signed already.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Link: https://lore.kernel.org/r/20231006102518.2452758-4-mingo@kernel.org
|
|
Use 'ilb_cpu' consistently in both functions.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Link: https://lore.kernel.org/r/20231006102518.2452758-3-mingo@kernel.org
|
|
- Fix incorrect/misleading comments,
- clarify some others,
- fix typos & grammar,
- and use more consistent style throughout.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Link: https://lore.kernel.org/r/20231006102518.2452758-2-mingo@kernel.org
|
|
The old pick_eevdf() could fail to find the actual earliest eligible
deadline when it descended to the right looking for min_deadline, but
it turned out that that min_deadline wasn't actually eligible. In that
case we need to go back and search through any left branches we
skipped looking for the actual best _eligible_ min_deadline.
This is more expensive, but still O(log n), and at worst should only
involve descending two branches of the rbtree.
I've run this through a userspace stress test (thank you
tools/lib/rbtree.c), so hopefully this implementation doesn't miss any
corner cases.
Fixes: 147f3efaa241 ("sched/fair: Implement an EEVDF-like scheduling policy")
Signed-off-by: Ben Segall <bsegall@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/xm261qego72d.fsf_-_@google.com
|
|
Marek and Biju reported instances of:
"EEVDF scheduling fail, picking leftmost"
which Mike correlated with cgroup scheduling and the min_deadline heap
getting corrupted; some trace output confirms:
> And yeah, min_deadline is hosed somehow:
>
> validate_cfs_rq: --- /
> __print_se: ffff88845cf48080 w: 1024 ve: -58857638 lag: 870381 vd: -55861854 vmd: -66302085 E (11372/tr)
> __print_se: ffff88810d165800 w: 25 ve: -80323686 lag: 22336429 vd: -41496434 vmd: -66302085 E (-1//autogroup-31)
> __print_se: ffff888108379000 w: 25 ve: 0 lag: -57987257 vd: 114632828 vmd: 114632828 N (-1//autogroup-33)
> validate_cfs_rq: min_deadline: -55861854 avg_vruntime: -62278313462 / 1074 = -57987256
Turns out that reweight_entity(), which tries really hard to be fast,
does not do the normal dequeue+update+enqueue pattern but *does* scale
the deadline.
However, it then fails to propagate the updated deadline value up the
heap.
Fixes: 147f3efaa241 ("sched/fair: Implement an EEVDF-like scheduling policy")
Reported-by: Marek Szyprowski <m.szyprowski@samsung.com>
Reported-by: Biju Das <biju.das.jz@bp.renesas.com>
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Marek Szyprowski <m.szyprowski@samsung.com>
Tested-by: Biju Das <biju.das.jz@bp.renesas.com>
Tested-by: Mike Galbraith <efault@gmx.de>
Link: https://lkml.kernel.org/r/20231006192445.GE743@noisy.programming.kicks-ass.net
|
|
the branch
Signed-off-by: Ingo Molnar <mingo@kernel.org>
|
|
The expectation is that placing a task at avg_vruntime() makes it
eligible. Turns out there is a corner case where this is not the case.
Specifically, avg_vruntime() relies on the fact that integer division
is a flooring function (eg. it discards the remainder). By this
property the value returned is slightly left of the true average.
However! when the average is a negative (relative to min_vruntime) the
effect is flipped and it becomes a ceil, with the result that the
returned value is just right of the average and thus not eligible.
Fixes: af4cf40470c2 ("sched/fair: Add cfs_rq::avg_vruntime")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
|
|
Tasks that never consume their full slice would not update their slice value.
This means that tasks that are spawned before the sysctl scaling keep their
original (UP) slice length.
Fixes: 147f3efaa241 ("sched/fair: Implement an EEVDF-like scheduling policy")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20230915124822.847197830@noisy.programming.kicks-ass.net
|
|
It was useful to track feec() placement decision and debug the spare
capacity and optimization issues vs uclamp_max.
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230916232955.2099394-4-qyousef@layalina.io
|
|
find_energy_efficient_cpu() bails out early if effective util of the
task is 0 as the delta at this point will be zero and there's nothing
for EAS to do. When uclamp is being used, this could lead to wrong
decisions when uclamp_max is set to 0. In this case the task is capped
to performance point 0, but it is actually running and consuming energy
and we can benefit from EAS energy calculations.
Rework the condition so that it bails out when both util and uclamp_min
are 0.
We can do that without needing to use uclamp_task_util(); remove it.
Fixes: d81304bc6193 ("sched/uclamp: Cater for uclamp in find_energy_efficient_cpu()'s early exit condition")
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230916232955.2099394-3-qyousef@layalina.io
|
|
When uclamp_max is being used, the util of the task could be higher than
the spare capacity of the CPU, but due to uclamp_max value we force-fit
it there.
The way the condition for checking for max_spare_cap in
find_energy_efficient_cpu() was constructed; it ignored any CPU that has
its spare_cap less than or _equal_ to max_spare_cap. Since we initialize
max_spare_cap to 0; this lead to never setting max_spare_cap_cpu and
hence ending up never performing compute_energy() for this cluster and
missing an opportunity for a better energy efficient placement to honour
uclamp_max setting.
max_spare_cap = 0;
cpu_cap = capacity_of(cpu) - cpu_util(p); // 0 if cpu_util(p) is high
...
util_fits_cpu(...); // will return true if uclamp_max forces it to fit
...
// this logic will fail to update max_spare_cap_cpu if cpu_cap is 0
if (cpu_cap > max_spare_cap) {
max_spare_cap = cpu_cap;
max_spare_cap_cpu = cpu;
}
prev_spare_cap suffers from a similar problem.
Fix the logic by converting the variables into long and treating -1
value as 'not populated' instead of 0 which is a viable and correct
spare capacity value. We need to be careful signed comparison is used
when comparing with cpu_cap in one of the conditions.
Fixes: 1d42509e475c ("sched/fair: Make EAS wakeup placement consider uclamp restrictions")
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20230916232955.2099394-2-qyousef@layalina.io
|
|
We've observed the following warning being hit in
distribute_cfs_runtime():
SCHED_WARN_ON(cfs_rq->runtime_remaining > 0)
We have the following race:
- CPU 0: running bandwidth distribution (distribute_cfs_runtime).
Inspects the local cfs_rq and makes its runtime_remaining positive.
However, we defer unthrottling the local cfs_rq until after
considering all remote cfs_rq's.
- CPU 1: starts running bandwidth distribution from the slack timer. When
it finds the cfs_rq for CPU 0 on the throttled list, it observers the
that the cfs_rq is throttled, yet is not on the CSD list, and has a
positive runtime_remaining, thus triggering the warning in
distribute_cfs_runtime.
To fix this, we can rework the local unthrottling logic to put the local
cfs_rq on a local list, so that any future bandwidth distributions will
realize that the cfs_rq is about to be unthrottled.
Signed-off-by: Josh Don <joshdon@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20230922230535.296350-2-joshdon@google.com
|
|
This makes the following patch cleaner by avoiding extra CONFIG_SMP
conditionals on the availability of rq->throttled_csd_list.
Signed-off-by: Josh Don <joshdon@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20230922230535.296350-1-joshdon@google.com
|
|
The /proc/sys/kernel/sched_child_runs_first knob is no longer connected since:
5e963f2bd4654 ("sched/fair: Commit to EEVDF")
Remove it.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20230920130025.412071-2-bigeasy@linutronix.de
|
|
The name is a bit opaque - make it clear that this is about wakeup
preemption.
Also rename the ->check_preempt_curr() methods similarly.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
|
|
Other scheduling classes already postfix their similar methods
with the class name.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
|
|
When using sysbench to benchmark Postgres in a single docker instance
with sysbench's nr_threads set to nr_cpu, it is observed there are times
update_cfs_group() and update_load_avg() shows noticeable overhead on
a 2sockets/112core/224cpu Intel Sapphire Rapids(SPR):
13.75% 13.74% [kernel.vmlinux] [k] update_cfs_group
10.63% 10.04% [kernel.vmlinux] [k] update_load_avg
Annotate shows the cycles are mostly spent on accessing tg->load_avg
with update_load_avg() being the write side and update_cfs_group() being
the read side. tg->load_avg is per task group and when different tasks
of the same taskgroup running on different CPUs frequently access
tg->load_avg, it can be heavily contended.
E.g. when running postgres_sysbench on a 2sockets/112cores/224cpus Intel
Sappire Rapids, during a 5s window, the wakeup number is 14millions and
migration number is 11millions and with each migration, the task's load
will transfer from src cfs_rq to target cfs_rq and each change involves
an update to tg->load_avg. Since the workload can trigger as many wakeups
and migrations, the access(both read and write) to tg->load_avg can be
unbound. As a result, the two mentioned functions showed noticeable
overhead. With netperf/nr_client=nr_cpu/UDP_RR, the problem is worse:
during a 5s window, wakeup number is 21millions and migration number is
14millions; update_cfs_group() costs ~25% and update_load_avg() costs ~16%.
Reduce the overhead by limiting updates to tg->load_avg to at most once
per ms. The update frequency is a tradeoff between tracking accuracy and
overhead. 1ms is chosen because PELT window is roughly 1ms and it
delivered good results for the tests that I've done. After this change,
the cost of accessing tg->load_avg is greatly reduced and performance
improved. Detailed test results below.
==============================
postgres_sysbench on SPR:
25%
base: 42382±19.8%
patch: 50174±9.5% (noise)
50%
base: 67626±1.3%
patch: 67365±3.1% (noise)
75%
base: 100216±1.2%
patch: 112470±0.1% +12.2%
100%
base: 93671±0.4%
patch: 113563±0.2% +21.2%
==============================
hackbench on ICL:
group=1
base: 114912±5.2%
patch: 117857±2.5% (noise)
group=4
base: 359902±1.6%
patch: 361685±2.7% (noise)
group=8
base: 461070±0.8%
patch: 491713±0.3% +6.6%
group=16
base: 309032±5.0%
patch: 378337±1.3% +22.4%
=============================
hackbench on SPR:
group=1
base: 100768±2.9%
patch: 103134±2.9% (noise)
group=4
base: 413830±12.5%
patch: 378660±16.6% (noise)
group=8
base: 436124±0.6%
patch: 490787±3.2% +12.5%
group=16
base: 457730±3.2%
patch: 680452±1.3% +48.8%
============================
netperf/udp_rr on ICL
25%
base: 114413±0.1%
patch: 115111±0.0% +0.6%
50%
base: 86803±0.5%
patch: 86611±0.0% (noise)
75%
base: 35959±5.3%
patch: 49801±0.6% +38.5%
100%
base: 61951±6.4%
patch: 70224±0.8% +13.4%
===========================
netperf/udp_rr on SPR
25%
base: 104954±1.3%
patch: 107312±2.8% (noise)
50%
base: 55394±4.6%
patch: 54940±7.4% (noise)
75%
base: 13779±3.1%
patch: 36105±1.1% +162%
100%
base: 9703±3.7%
patch: 28011±0.2% +189%
==============================================
netperf/tcp_stream on ICL (all in noise range)
25%
base: 43092±0.1%
patch: 42891±0.5%
50%
base: 19278±14.9%
patch: 22369±7.2%
75%
base: 16822±3.0%
patch: 17086±2.3%
100%
base: 18216±0.6%
patch: 18078±2.9%
===============================================
netperf/tcp_stream on SPR (all in noise range)
25%
base: 34491±0.3%
patch: 34886±0.5%
50%
base: 19278±14.9%
patch: 22369±7.2%
75%
base: 16822±3.0%
patch: 17086±2.3%
100%
base: 18216±0.6%
patch: 18078±2.9%
Reported-by: Nitin Tekchandani <nitin.tekchandani@intel.com>
Suggested-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Aaron Lu <aaron.lu@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Reviewed-by: David Vernet <void@manifault.com>
Tested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Tested-by: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
Link: https://lkml.kernel.org/r/20230912065808.2530-2-aaron.lu@intel.com
|
|
We don't need to maintain per-queue leaf_cfs_rq_list on !SMP, since
it's used for cfs_rq load tracking & balancing on SMP.
But sched debug interface uses it to print per-cfs_rq stats.
This patch fixes the !SMP version of cfs_rq_is_decayed(), so the
per-queue leaf_cfs_rq_list is also maintained correctly on !SMP,
to fix the warning in assert_list_leaf_cfs_rq().
Fixes: 0a00a354644e ("sched/fair: Delete useless condition in tg_unthrottle_up()")
Reported-by: Leo Yu-Chi Liang <ycliang@andestech.com>
Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Leo Yu-Chi Liang <ycliang@andestech.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Closes: https://lore.kernel.org/all/ZN87UsqkWcFLDxea@swlinux02/
Link: https://lore.kernel.org/r/20230913132031.2242151-1-chengming.zhou@linux.dev
|
|
task_numa_placement() searches for a nearest node to migrate by calling
for_each_node_state(). Now that we have numa_nearest_node(), switch to
using it.
Signed-off-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Link: https://lore.kernel.org/r/20230819141239.287290-3-yury.norov@gmail.com
|
|
For SMT4, any group with more than 2 tasks will be marked as
group_smt_balance. Retain the behaviour of group_has_spare by marking
the busiest group as the group which has the least number of idle_cpus.
Also, handle rounding effect of adding (ncores_local + ncores_busy) when
the local is fully idle and busy group imbalance is less than 2 tasks.
Local group should try to pull at least 1 task in this case so imbalance
should be set to 2 instead.
Fixes: fee1759e4f04 ("sched/fair: Determine active load balance for SMT sched groups")
Acked-by: Shrikanth Hegde <sshegde@linux.vnet.ibm.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: http://lkml.kernel.org/r/6cd1633036bb6b651af575c32c2a9608a106702c.camel@linux.intel.com
|
|
should_we_balance() is called in load_balance() to find out if the CPU that
is trying to do the load balance is the right one or not.
With commit:
b1bfeab9b002("sched/fair: Consider the idle state of the whole core for load balance")
the code tries to find an idle core to do the load balancing
and falls back on an idle sibling CPU if there is no idle core.
However, on larger SMT systems, it could be needlessly iterating to find a
idle by scanning all the CPUs in an non-idle core. If the core is not idle,
and first SMT sibling which is idle has been found, then its not needed to
check other SMT siblings for idleness
Lets say in SMT4, Core0 has 0,2,4,6 and CPU0 is BUSY and rest are IDLE.
balancing domain is MC/DIE. CPU2 will be set as the first idle_smt and
same process would be repeated for CPU4 and CPU6 but this is unnecessary.
Since calling is_core_idle loops through all CPU's in the SMT mask, effect
is multiplied by weight of smt_mask. For example,when say 1 CPU is busy,
we would skip loop for 2 CPU's and skip iterating over 8CPU's. That
effect would be more in DIE/NUMA domain where there are more cores.
Testing and performance evaluation
==================================
The test has been done on this system which has 12 cores, i.e 24 small
cores with SMT=4:
lscpu
Architecture: ppc64le
Byte Order: Little Endian
CPU(s): 96
On-line CPU(s) list: 0-95
Model name: POWER10 (architected), altivec supported
Thread(s) per core: 8
Used funclatency bcc tool to evaluate the time taken by should_we_balance(). For
base tip/sched/core the time taken is collected by making the
should_we_balance() noinline. time is in nanoseconds. The values are
collected by running the funclatency tracer for 60 seconds. values are
average of 3 such runs. This represents the expected reduced time with
patch.
tip/sched/core was at commit:
2f88c8e802c8 ("sched/eevdf/doc: Modify the documented knob to base_slice_ns as well")
Results:
------------------------------------------------------------------------------
workload tip/sched/core with_patch(%gain)
------------------------------------------------------------------------------
idle system 809.3 695.0(16.45)
stress ng – 12 threads -l 100 1013.5 893.1(13.49)
stress ng – 24 threads -l 100 1073.5 980.0(9.54)
stress ng – 48 threads -l 100 683.0 641.0(6.55)
stress ng – 96 threads -l 100 2421.0 2300(5.26)
stress ng – 96 threads -l 15 375.5 377.5(-0.53)
stress ng – 96 threads -l 25 635.5 637.5(-0.31)
stress ng – 96 threads -l 35 934.0 891.0(4.83)
Ran schbench(old), hackbench and stress_ng to evaluate the workload
performance between tip/sched/core and with patch.
No modification to tip/sched/core
TL;DR:
Good improvement is seen with schbench. when hackbench and stress_ng
runs for longer good improvement is seen.
------------------------------------------------------------------------------
schbench(old) tip +patch(%gain)
10 iterations sched/core
------------------------------------------------------------------------------
1 Threads
50.0th: 8.00 9.00(-12.50)
75.0th: 9.60 9.00(6.25)
90.0th: 11.80 10.20(13.56)
95.0th: 12.60 10.40(17.46)
99.0th: 13.60 11.90(12.50)
99.5th: 14.10 12.60(10.64)
99.9th: 15.90 14.60(8.18)
2 Threads
50.0th: 9.90 9.20(7.07)
75.0th: 12.60 10.10(19.84)
90.0th: 15.50 12.00(22.58)
95.0th: 17.70 14.00(20.90)
99.0th: 21.20 16.90(20.28)
99.5th: 22.60 17.50(22.57)
99.9th: 30.40 19.40(36.18)
4 Threads
50.0th: 12.50 10.60(15.20)
75.0th: 15.30 12.00(21.57)
90.0th: 18.60 14.10(24.19)
95.0th: 21.30 16.20(23.94)
99.0th: 26.00 20.70(20.38)
99.5th: 27.60 22.50(18.48)
99.9th: 33.90 31.40(7.37)
8 Threads
50.0th: 16.30 14.30(12.27)
75.0th: 20.20 17.40(13.86)
90.0th: 24.50 21.90(10.61)
95.0th: 27.30 24.70(9.52)
99.0th: 35.00 31.20(10.86)
99.5th: 46.40 33.30(28.23)
99.9th: 89.30 57.50(35.61)
16 Threads
50.0th: 22.70 20.70(8.81)
75.0th: 30.10 27.40(8.97)
90.0th: 36.00 32.80(8.89)
95.0th: 39.60 36.40(8.08)
99.0th: 49.20 44.10(10.37)
99.5th: 64.90 50.50(22.19)
99.9th: 143.50 100.60(29.90)
32 Threads
50.0th: 34.60 35.50(-2.60)
75.0th: 48.20 50.50(-4.77)
90.0th: 59.20 62.40(-5.41)
95.0th: 65.20 69.00(-5.83)
99.0th: 80.40 83.80(-4.23)
99.5th: 102.10 98.90(3.13)
99.9th: 727.10 506.80(30.30)
schbench does improve in general. There is some run to run variation with
schbench. Did a validation run to confirm that trend is similar.
------------------------------------------------------------------------------
hackbench tip +patch(%gain)
20 iterations, 50000 loops sched/core
------------------------------------------------------------------------------
Process 10 groups : 11.74 11.70(0.34)
Process 20 groups : 22.73 22.69(0.18)
Process 30 groups : 33.39 33.40(-0.03)
Process 40 groups : 43.73 43.61(0.27)
Process 50 groups : 53.82 54.35(-0.98)
Process 60 groups : 64.16 65.29(-1.76)
thread 10 Time : 12.81 12.79(0.16)
thread 20 Time : 24.63 24.47(0.65)
Process(Pipe) 10 Time : 6.40 6.34(0.94)
Process(Pipe) 20 Time : 10.62 10.63(-0.09)
Process(Pipe) 30 Time : 15.09 14.84(1.66)
Process(Pipe) 40 Time : 19.42 19.01(2.11)
Process(Pipe) 50 Time : 24.04 23.34(2.91)
Process(Pipe) 60 Time : 28.94 27.51(4.94)
thread(Pipe) 10 Time : 6.96 6.87(1.29)
thread(Pipe) 20 Time : 11.74 11.73(0.09)
hackbench shows slight improvement with pipe. Slight degradation in process.
------------------------------------------------------------------------------
stress_ng tip +patch(%gain)
10 iterations 100000 cpu_ops sched/core
------------------------------------------------------------------------------
--cpu=96 -util=100 Time taken : 5.30, 5.01(5.47)
--cpu=48 -util=100 Time taken : 7.94, 6.73(15.24)
--cpu=24 -util=100 Time taken : 11.67, 8.75(25.02)
--cpu=12 -util=100 Time taken : 15.71, 15.02(4.39)
--cpu=96 -util=10 Time taken : 22.71, 22.19(2.29)
--cpu=96 -util=20 Time taken : 12.14, 12.37(-1.89)
--cpu=96 -util=30 Time taken : 8.76, 8.86(-1.14)
--cpu=96 -util=40 Time taken : 7.13, 7.14(-0.14)
--cpu=96 -util=50 Time taken : 6.10, 6.13(-0.49)
--cpu=96 -util=60 Time taken : 5.42, 5.41(0.18)
--cpu=96 -util=70 Time taken : 4.94, 4.94(0.00)
--cpu=96 -util=80 Time taken : 4.56, 4.53(0.66)
--cpu=96 -util=90 Time taken : 4.27, 4.26(0.23)
Good improvement seen with 24 CPUs. In this case only one CPU is busy,
and no core is idle. Decent improvement with 100% utilization case. no
difference in other utilization.
Fixes: b1bfeab9b002 ("sched/fair: Consider the idle state of the whole core for load balance")
Signed-off-by: Shrikanth Hegde <sshegde@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20230902081204.232218-1-sshegde@linux.vnet.ibm.com
|
|
The function update_entity_lag() is only used inside the kernel/sched/fair.c file.
Make it static.
Signed-off-by: Hao Jia <jiahao.os@bytedance.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20230829030325.69128-1-jiahao.os@bytedance.com
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull misc x86 cleanups from Ingo Molnar:
"The following commit deserves special mention:
22dc02f81cddd Revert "sched/fair: Move unused stub functions to header"
This is in x86/cleanups, because the revert is a re-application of a
number of cleanups that got removed inadvertedly"
[ This also effectively undoes the amd_check_microcode() microcode
declaration change I had done in my microcode loader merge in commit
42a7f6e3ffe0 ("Merge tag 'x86_microcode_for_v6.6_rc1' [...]").
I picked the declaration change by Arnd from this branch instead,
which put it in <asm/processor.h> instead of <asm/microcode.h> like I
had done in my merge resolution - Linus ]
* tag 'x86-cleanups-2023-08-28' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/platform/uv: Refactor code using deprecated strncpy() interface to use strscpy()
x86/hpet: Refactor code using deprecated strncpy() interface to use strscpy()
x86/platform/uv: Refactor code using deprecated strcpy()/strncpy() interfaces to use strscpy()
x86/qspinlock-paravirt: Fix missing-prototype warning
x86/paravirt: Silence unused native_pv_lock_init() function warning
x86/alternative: Add a __alt_reloc_selftest() prototype
x86/purgatory: Include header for warn() declaration
x86/asm: Avoid unneeded __div64_32 function definition
Revert "sched/fair: Move unused stub functions to header"
x86/apic: Hide unused safe_smp_processor_id() on 32-bit UP
x86/cpu: Fix amd_check_microcode() declaration
|
