Skip to content

Commit

Permalink
runtime: don't usleep() in runqgrab()
Browse files Browse the repository at this point in the history 
The most recently goready()'d G on each P is given a special position in
the P's runqueue, p.runnext. Other Ps steal p.runnext only as a last
resort, and usleep(3) before doing so: findRunnable() => stealWork() =>
runqsteal() => runqgrab(). As documented in runqgrab(), this is to
reduce thrashing of Gs between Ps in cases where one goroutine wakes another
and then "almost immediately" blocks.

On Linux, usleep() is implemented by invoking the nanosleep system call.
Syscall timeouts in the Linux kernel are subject to timer slack, as
documented by the man page for syscall prctl, section
"PR_SET_TIMERSLACK". Experimentally, short timeouts can expect to expire
50 microseconds late regardless of other system activity. Thus, on
Linux, usleep(3) typically sleeps for at least 53 microseconds, more
than 17x longer than intended.

A P must be in the spinning state in order to attempt work-stealing.
While at least one P is spinning, wakep() will refuse to wake a new
spinning P. One P sleeping in runqgrab() thus prevents further threads
from being woken in response to e.g. goroutine wakeups *globally*
(throughout the process). Futex wake-to-wakeup latency is approximately
20 microseconds, so sleeping for 53 microseconds can significantly
increase goroutine wakeup latency by delaying thread wakeup.

Fix this by timestamping Gs when they are runqput() into p.runnext, and
causing runqgrab() to indicate to findRunnable() that it should loop if
p.runnext is not yet stealable.

Alternative fixes considered:

- osyield() on Linux as we do on a few other platforms. On Linux,
  osyield() is implemented by the sched_yield system call, which IIUC
  causes the calling thread to yield its timeslice to any thread on its
  runqueue that it would not preempt on wakeup, potentially introducing
  even larger latencies on busy systems. See also
  https://www.realworldtech.com/forum/?threadid=189711&curpostid=189752
  for a case against sched_yield on semantic grounds.

- Replace the usleep() with a spin loop in-place. This tends to waste
  the spinning P's time, since it can't check other runqueues and the
  number of calls to runqgrab() - and therefore sleeps -  is linear in
  the number of Ps. Empirically, it introduces regressions not observed
  in this change.

- Change thread timer slack using prctl(PR_SET_TIMERSLACK). In practice,
  user programs will have been tuned based on the default timer slack
  value, so tampering with this may introduce regressions into existing
  programs.

Unfortunately, this is a load-bearing bug. In programs with goroutines
that frequently wake up goroutines and then immediately block, this bug
significantly reduces overhead from useless thread wakeups in wakep().
In golang.org/x/benchmarks, this manifests most clearly as regressions
in benchmark dustin_broadcast. To avoid this regression, we need to
intentionally throttle wakep() => acquirem().

Thus, this change also introduces a "need-wakep()" prediction mechanism,
which causes goready() and newproc() to call wakep() only if the calling
goroutine is predicted not to immediately block. To handle
mispredictions, sysmon is changed to wakep() if it detects
underutilization. The current prediction algorithm is simple, but
appears to be effective; it can be improved in the future as warranted.

Results from golang.org/x/benchmarks:
(Baseline is go1.20.1; experiment is go1.20.1 plus this change)

shortname: ajstarks_deck_generate
goos: linux
goarch: amd64
pkg: github.com/ajstarks/deck/generate
cpu: Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz
           │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
           │                    sec/op                    │         sec/op           vs base               │
Arc-12                                        3.857µ ± 5%               3.753µ ± 5%       ~ (p=0.424 n=10)
Polygon-12                                    7.074µ ± 6%               6.969µ ± 4%       ~ (p=0.190 n=10)
geomean                                       5.224µ                    5.114µ       -2.10%

shortname: aws_jsonutil
pkg: github.com/aws/aws-sdk-go/private/protocol/json/jsonutil
              │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
              │                    sec/op                    │         sec/op           vs base               │
BuildJSON-12                                     5.602µ ± 3%               5.600µ ± 2%       ~ (p=0.896 n=10)
StdlibJSON-12                                    3.843µ ± 2%               3.828µ ± 2%       ~ (p=0.224 n=10)
geomean                                          4.640µ                    4.630µ       -0.22%

shortname: benhoyt_goawk_1_18
pkg: github.com/benhoyt/goawk/interp
                       │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                       │                    sec/op                    │         sec/op           vs base               │
RecursiveFunc-12                                          17.79µ ± 3%               17.65µ ± 3%       ~ (p=0.436 n=10)
RegexMatch-12                                             815.8n ± 4%               823.3n ± 1%       ~ (p=0.353 n=10)
RepeatExecProgram-12                                      21.30µ ± 6%               21.69µ ± 3%       ~ (p=0.052 n=10)
RepeatNew-12                                              79.21n ± 4%               79.73n ± 3%       ~ (p=0.529 n=10)
RepeatIOExecProgram-12                                    41.83µ ± 1%               42.07µ ± 2%       ~ (p=0.796 n=10)
RepeatIONew-12                                            1.195µ ± 3%               1.196µ ± 2%       ~ (p=1.000 n=10)
geomean                                                   3.271µ                    3.288µ       +0.54%

shortname: bindata
pkg: github.com/kevinburke/go-bindata
           │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
           │                    sec/op                    │            sec/op             vs base          │
Bindata-12                                    316.2m ± 5%                    309.7m ± 4%  ~ (p=0.436 n=10)

           │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
           │                     B/s                      │             B/s               vs base          │
Bindata-12                                   20.71Mi ± 5%                   21.14Mi ± 4%  ~ (p=0.436 n=10)

           │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
           │                     B/op                     │             B/op              vs base          │
Bindata-12                                   183.0Mi ± 0%                   183.0Mi ± 0%  ~ (p=0.353 n=10)

           │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
           │                  allocs/op                   │          allocs/op            vs base          │
Bindata-12                                    5.790k ± 0%                    5.789k ± 0%  ~ (p=0.358 n=10)

shortname: bloom_bloom
pkg: github.com/bits-and-blooms/bloom/v3
                      │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                      │                    sec/op                    │         sec/op           vs base               │
SeparateTestAndAdd-12                                    414.6n ± 4%               413.9n ± 2%       ~ (p=0.895 n=10)
CombinedTestAndAdd-12                                    425.8n ± 9%               419.8n ± 8%       ~ (p=0.353 n=10)
geomean                                                  420.2n                    416.9n       -0.78%

shortname: capnproto2
pkg: zombiezen.com/go/capnproto2
                               │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                               │                    sec/op                    │         sec/op           vs base               │
TextMovementBetweenSegments-12                                    320.5µ ± 5%              318.4µ ± 10%       ~ (p=0.579 n=10)
Growth_MultiSegment-12                                            13.63m ± 1%              13.87m ±  2%  +1.71% (p=0.029 n=10)
geomean                                                           2.090m                   2.101m        +0.52%

                       │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                       │                     B/s                      │           B/s            vs base               │
Growth_MultiSegment-12                                   73.35Mi ± 1%              72.12Mi ± 2%  -1.68% (p=0.027 n=10)

                       │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                       │                     B/op                     │             B/op              vs base          │
Growth_MultiSegment-12                                   1.572Mi ± 0%                   1.572Mi ± 0%  ~ (p=0.320 n=10)

                       │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                       │                  allocs/op                   │         allocs/op           vs base            │
Growth_MultiSegment-12                                     21.00 ± 0%                   21.00 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

shortname: cespare_mph
pkg: github.com/cespare/mph
         │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
         │                    sec/op                    │            sec/op             vs base          │
Build-12                                    32.72m ± 2%                    32.49m ± 1%  ~ (p=0.280 n=10)

shortname: commonmark_markdown
pkg: gitlab.com/golang-commonmark/markdown
                          │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                          │                    sec/op                    │         sec/op           vs base               │
RenderSpecNoHTML-12                                          10.09m ± 2%               10.18m ± 3%       ~ (p=0.796 n=10)
RenderSpec-12                                                10.19m ± 1%               10.11m ± 3%       ~ (p=0.684 n=10)
RenderSpecBlackFriday2-12                                    6.793m ± 5%               6.946m ± 2%       ~ (p=0.063 n=10)
geomean                                                      8.872m                    8.944m       +0.81%

shortname: dustin_broadcast
pkg: github.com/dustin/go-broadcast
                      │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                      │                    sec/op                    │         sec/op          vs base                │
DirectSend-12                                            570.5n ± 7%              355.2n ± 2%  -37.74% (p=0.000 n=10)
ParallelDirectSend-12                                    549.0n ± 5%              360.9n ± 3%  -34.25% (p=0.000 n=10)
ParallelBrodcast-12                                      788.7n ± 2%              486.0n ± 4%  -38.37% (p=0.000 n=10)
MuxBrodcast-12                                           788.6n ± 4%              471.5n ± 6%  -40.21% (p=0.000 n=10)
geomean                                                  664.4n                   414.0n       -37.68%

shortname: dustin_humanize
pkg: github.com/dustin/go-humanize
                 │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                 │                    sec/op                    │            sec/op             vs base          │
ParseBigBytes-12                                    1.964µ ± 5%                    1.941µ ± 3%  ~ (p=0.289 n=10)

shortname: ericlagergren_decimal
pkg: github.com/ericlagergren/decimal/benchmarks
                                       │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                       │                    sec/op                    │         sec/op           vs base               │
Pi/foo=ericlagergren_(Go)/prec=100-12                                     147.5µ ± 2%               147.5µ ± 1%       ~ (p=0.912 n=10)
Pi/foo=ericlagergren_(GDA)/prec=100-12                                    329.6µ ± 1%               332.1µ ± 2%       ~ (p=0.063 n=10)
Pi/foo=shopspring/prec=100-12                                             680.5µ ± 4%               688.6µ ± 2%       ~ (p=0.481 n=10)
Pi/foo=apmckinlay/prec=100-12                                             2.541µ ± 4%               2.525µ ± 3%       ~ (p=0.218 n=10)
Pi/foo=go-inf/prec=100-12                                                 169.5µ ± 3%               170.7µ ± 3%       ~ (p=0.218 n=10)
Pi/foo=float64/prec=100-12                                                4.136µ ± 3%               4.162µ ± 6%       ~ (p=0.436 n=10)
geomean                                                                   62.38µ                    62.66µ       +0.45%

shortname: ethereum_bitutil
pkg: github.com/ethereum/go-ethereum/common/bitutil
                         │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                         │                    sec/op                    │         sec/op          vs base                │
FastTest2KB-12                                              130.4n ± 1%              131.5n ± 1%        ~ (p=0.093 n=10)
BaseTest2KB-12                                              624.8n ± 2%              983.0n ± 2%  +57.32% (p=0.000 n=10)
Encoding4KBVerySparse-12                                    21.48µ ± 3%              22.20µ ± 3%   +3.37% (p=0.005 n=10)
geomean                                                     1.205µ                   1.421µ       +17.94%

                         │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                         │                     B/op                     │            B/op             vs base            │
Encoding4KBVerySparse-12                                   9.750Ki ± 0%                 9.750Ki ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

                         │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                         │                  allocs/op                   │         allocs/op           vs base            │
Encoding4KBVerySparse-12                                     15.00 ± 0%                   15.00 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

shortname: ethereum_core
pkg: github.com/ethereum/go-ethereum/core
                             │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                             │                    sec/op                    │         sec/op           vs base               │
PendingDemotion10000-12                                         96.72n ± 4%               98.55n ± 2%       ~ (p=0.055 n=10)
FuturePromotion10000-12                                         2.128n ± 3%               2.093n ± 3%       ~ (p=0.896 n=10)
PoolBatchInsert10000-12                                         642.6m ± 2%               642.1m ± 5%       ~ (p=0.796 n=10)
PoolBatchLocalInsert10000-12                                    805.2m ± 2%               826.6m ± 4%       ~ (p=0.105 n=10)
geomean                                                         101.6µ                    102.3µ       +0.69%

shortname: ethereum_corevm
pkg: github.com/ethereum/go-ethereum/core/vm
            │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
            │                    sec/op                    │         sec/op           vs base               │
OpDiv128-12                                    137.4n ± 3%               139.5n ± 1%  +1.56% (p=0.024 n=10)

shortname: ethereum_ecies
pkg: github.com/ethereum/go-ethereum/crypto/ecies
                    │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                    │                    sec/op                    │         sec/op           vs base               │
GenerateKeyP256-12                                     15.67µ ± 6%               15.66µ ± 3%       ~ (p=0.971 n=10)
GenSharedKeyP256-12                                    51.09µ ± 6%               52.09µ ± 4%       ~ (p=0.631 n=10)
GenSharedKeyS256-12                                    47.24µ ± 2%               46.67µ ± 3%       ~ (p=0.247 n=10)
geomean                                                33.57µ                    33.64µ       +0.21%

shortname: ethereum_ethash
pkg: github.com/ethereum/go-ethereum/consensus/ethash
                  │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                  │                    sec/op                    │            sec/op             vs base          │
HashimotoLight-12                                    1.116m ± 5%                    1.112m ± 2%  ~ (p=0.684 n=10)

shortname: ethereum_trie
pkg: github.com/ethereum/go-ethereum/trie
                                │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                │                    sec/op                    │         sec/op           vs base               │
HashFixedSize/10K-12                                               9.236m ± 1%               9.106m ± 1%  -1.40% (p=0.019 n=10)
CommitAfterHashFixedSize/10K-12                                    19.60m ± 1%               19.51m ± 1%       ~ (p=0.796 n=10)
geomean                                                            13.45m                    13.33m       -0.93%

                                │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                │                     B/op                     │          B/op            vs base               │
HashFixedSize/10K-12                                              6.036Mi ± 0%              6.037Mi ± 0%       ~ (p=0.247 n=10)
CommitAfterHashFixedSize/10K-12                                   8.626Mi ± 0%              8.626Mi ± 0%       ~ (p=0.280 n=10)
geomean                                                           7.216Mi                   7.216Mi       +0.01%

                                │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                │                  allocs/op                   │        allocs/op         vs base               │
HashFixedSize/10K-12                                               77.17k ± 0%               77.17k ± 0%       ~ (p=0.050 n=10)
CommitAfterHashFixedSize/10K-12                                    79.99k ± 0%               79.99k ± 0%       ~ (p=0.391 n=10)
geomean                                                            78.56k                    78.57k       +0.00%

shortname: gonum_blas_native
pkg: gonum.org/v1/gonum/blas/gonum
                         │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                         │                    sec/op                    │         sec/op           vs base               │
Dnrm2MediumPosInc-12                                        1.953µ ± 2%               1.940µ ± 5%       ~ (p=0.989 n=10)
DasumMediumUnitaryInc-12                                    932.5n ± 1%               931.2n ± 1%       ~ (p=0.753 n=10)
geomean                                                     1.349µ                    1.344µ       -0.40%

shortname: gonum_community
pkg: gonum.org/v1/gonum/graph/community
                            │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                            │                    sec/op                    │            sec/op             vs base          │
LouvainDirectedMultiplex-12                                    26.40m ± 1%                    26.64m ± 1%  ~ (p=0.165 n=10)

shortname: gonum_lapack_native
pkg: gonum.org/v1/gonum/lapack/gonum
                      │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                      │                    sec/op                    │         sec/op           vs base               │
Dgeev/Circulant10-12                                     41.97µ ± 6%               42.90µ ± 4%       ~ (p=0.143 n=10)
Dgeev/Circulant100-12                                    12.13m ± 4%               12.30m ± 3%       ~ (p=0.796 n=10)
geomean                                                  713.4µ                    726.4µ       +1.81%

shortname: gonum_mat
pkg: gonum.org/v1/gonum/mat
                                  │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                  │                    sec/op                    │         sec/op           vs base               │
MulWorkspaceDense1000Hundredth-12                                   89.78m ±  0%              81.48m ±  1%  -9.24% (p=0.000 n=10)
ScaleVec10000Inc20-12                                               7.204µ ± 36%              8.450µ ± 35%       ~ (p=0.853 n=10)
geomean                                                             804.2µ                    829.7µ        +3.18%

shortname: gonum_topo
pkg: gonum.org/v1/gonum/graph/topo
                          │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                          │                    sec/op                    │         sec/op           vs base               │
TarjanSCCGnp_10_tenth-12                                     7.251µ ± 1%               7.187µ ± 1%  -0.88% (p=0.025 n=10)
TarjanSCCGnp_1000_half-12                                    74.48m ± 2%               74.37m ± 4%       ~ (p=0.796 n=10)
geomean                                                      734.8µ                    731.1µ       -0.51%

shortname: gonum_traverse
pkg: gonum.org/v1/gonum/graph/traverse
                                     │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                     │                    sec/op                    │         sec/op           vs base               │
WalkAllBreadthFirstGnp_10_tenth-12                                      3.517µ ± 1%               3.534µ ± 1%       ~ (p=0.343 n=10)
WalkAllBreadthFirstGnp_1000_tenth-12                                    11.12m ± 6%               11.19m ± 2%       ~ (p=0.631 n=10)
geomean                                                                 197.8µ                    198.9µ       +0.54%

shortname: gtank_blake2s
pkg: github.com/gtank/blake2s
          │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
          │                    sec/op                    │            sec/op             vs base          │
Hash8K-12                                    18.96µ ± 4%                    18.82µ ± 5%  ~ (p=0.579 n=10)

          │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
          │                     B/s                      │             B/s               vs base          │
Hash8K-12                                   412.2Mi ± 4%                   415.2Mi ± 5%  ~ (p=0.579 n=10)

shortname: hugo_hugolib
pkg: github.com/gohugoio/hugo/hugolib
                            │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                            │                    sec/op                    │         sec/op          vs base                │
MergeByLanguage-12                                             529.9n ± 1%              531.5n ± 2%        ~ (p=0.305 n=10)
ResourceChainPostProcess-12                                    62.76m ± 3%              56.23m ± 2%  -10.39% (p=0.000 n=10)
ReplaceShortcodeTokens-12                                      2.727µ ± 3%              2.701µ ± 7%        ~ (p=0.592 n=10)
geomean                                                        44.92µ                   43.22µ        -3.80%

shortname: k8s_cache
pkg: k8s.io/client-go/tools/cache
                           │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                           │                    sec/op                    │         sec/op           vs base               │
Listener-12                                                   1.312µ ± 1%               1.199µ ± 1%  -8.62% (p=0.000 n=10)
ReflectorResyncChanMany-12                                    785.7n ± 4%               796.3n ± 3%       ~ (p=0.089 n=10)
geomean                                                       1.015µ                    976.9n       -3.76%

            │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
            │                     B/op                     │            B/op             vs base            │
Listener-12                                     16.00 ± 0%                   16.00 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

            │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
            │                  allocs/op                   │         allocs/op           vs base            │
Listener-12                                     1.000 ± 0%                   1.000 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

shortname: k8s_workqueue
pkg: k8s.io/client-go/util/workqueue
                                                         │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                                         │                    sec/op                    │         sec/op          vs base                │
ParallelizeUntil/pieces:1000,workers:10,chunkSize:1-12                                      244.6µ ± 1%              245.9µ ± 0%   +0.55% (p=0.023 n=10)
ParallelizeUntil/pieces:1000,workers:10,chunkSize:10-12                                     75.09µ ± 1%              63.54µ ± 1%  -15.37% (p=0.000 n=10)
ParallelizeUntil/pieces:1000,workers:10,chunkSize:100-12                                    49.47µ ± 2%              42.45µ ± 2%  -14.19% (p=0.000 n=10)
ParallelizeUntil/pieces:999,workers:10,chunkSize:13-12                                      68.51µ ± 1%              55.07µ ± 1%  -19.63% (p=0.000 n=10)
geomean                                                                                     88.82µ                   77.74µ       -12.47%

shortname: kanzi
pkg: github.com/flanglet/kanzi-go/benchmark
        │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
        │                    sec/op                    │         sec/op           vs base               │
BWTS-12                                   0.4479n ± 6%              0.4385n ± 7%       ~ (p=0.529 n=10)
FPAQ-12                                    17.03m ± 3%               17.42m ± 3%       ~ (p=0.123 n=10)
LZ-12                                      1.897m ± 2%               1.887m ± 4%       ~ (p=1.000 n=10)
MTFT-12                                    771.2µ ± 4%               785.8µ ± 3%       ~ (p=0.247 n=10)
geomean                                    57.79µ                    58.01µ       +0.38%

shortname: minio
pkg: github.com/minio/minio/cmd
                                │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                │                    sec/op                    │         sec/op          vs base                │
DecodehealingTracker-12                                            852.8n ± 5%              866.8n ± 5%        ~ (p=0.190 n=10)
AppendMsgReplicateDecision-12                                     0.5383n ± 4%             0.7598n ± 3%  +41.13% (p=0.000 n=10)
AppendMsgResyncTargetsInfo-12                                      4.785n ± 2%              4.639n ± 3%   -3.06% (p=0.003 n=10)
DataUpdateTracker-12                                               3.122µ ± 2%              1.880µ ± 3%  -39.77% (p=0.000 n=10)
MarshalMsgdataUsageCacheInfo-12                                    110.9n ± 2%              109.4n ± 3%        ~ (p=0.101 n=10)
geomean                                                            59.74n                   57.50n        -3.75%

                              │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                              │                     B/s                      │          B/s            vs base                │
DecodehealingTracker-12                                         347.8Mi ± 5%             342.2Mi ± 6%        ~ (p=0.190 n=10)
AppendMsgReplicateDecision-12                                   1.730Gi ± 3%             1.226Gi ± 3%  -29.14% (p=0.000 n=10)
AppendMsgResyncTargetsInfo-12                                   1.946Gi ± 2%             2.008Gi ± 3%   +3.15% (p=0.003 n=10)
DataUpdateTracker-12                                            312.5Ki ± 3%             517.6Ki ± 2%  +65.62% (p=0.000 n=10)
geomean                                                         139.1Mi                  145.4Mi        +4.47%

                                │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                │                     B/op                     │         B/op           vs base                 │
DecodehealingTracker-12                                           0.000 ± 0%                0.000 ± 0%       ~ (p=1.000 n=10) ¹
AppendMsgReplicateDecision-12                                     0.000 ± 0%                0.000 ± 0%       ~ (p=1.000 n=10) ¹
AppendMsgResyncTargetsInfo-12                                     0.000 ± 0%                0.000 ± 0%       ~ (p=1.000 n=10) ¹
DataUpdateTracker-12                                              340.0 ± 0%                339.0 ± 1%       ~ (p=0.737 n=10)
MarshalMsgdataUsageCacheInfo-12                                   96.00 ± 0%                96.00 ± 0%       ~ (p=1.000 n=10) ¹
geomean                                                                      ²                          -0.06%                ²
¹ all samples are equal
² summaries must be >0 to compute geomean

                                │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                │                  allocs/op                   │       allocs/op        vs base                 │
DecodehealingTracker-12                                           0.000 ± 0%                0.000 ± 0%       ~ (p=1.000 n=10) ¹
AppendMsgReplicateDecision-12                                     0.000 ± 0%                0.000 ± 0%       ~ (p=1.000 n=10) ¹
AppendMsgResyncTargetsInfo-12                                     0.000 ± 0%                0.000 ± 0%       ~ (p=1.000 n=10) ¹
DataUpdateTracker-12                                              9.000 ± 0%                9.000 ± 0%       ~ (p=1.000 n=10) ¹
MarshalMsgdataUsageCacheInfo-12                                   1.000 ± 0%                1.000 ± 0%       ~ (p=1.000 n=10) ¹
geomean                                                                      ²                          +0.00%                ²
¹ all samples are equal
² summaries must be >0 to compute geomean

shortname: semver
pkg: github.com/Masterminds/semver
                            │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                            │                    sec/op                    │            sec/op             vs base          │
ValidateVersionTildeFail-12                                    854.7n ± 2%                    842.7n ± 2%  ~ (p=0.123 n=10)

shortname: shopify_sarama
pkg: github.com/Shopify/sarama
                          │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                          │                    sec/op                    │         sec/op           vs base               │
Broker_Open-12                                               212.2µ ± 1%               205.9µ ± 2%  -2.95% (p=0.000 n=10)
Broker_No_Metrics_Open-12                                    132.9µ ± 1%               121.3µ ± 2%  -8.68% (p=0.000 n=10)
geomean                                                      167.9µ                    158.1µ       -5.86%

shortname: spexs2
pkg: github.com/egonelbre/spexs2/_benchmark
              │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
              │                    sec/op                    │         sec/op           vs base               │
Run/10k/1-12                                      23.29 ± 1%                23.11 ± 2%       ~ (p=0.315 n=10)
Run/10k/16-12                                     5.648 ± 2%                5.462 ± 4%  -3.30% (p=0.004 n=10)
geomean                                           11.47                     11.23       -2.06%

shortname: sweet-biogo-igor
goos:
goarch:
pkg:
cpu:
          │ ./sweet/results/biogo-igor/baseline.results │ ./sweet/results/biogo-igor/experiment.results │
          │                   sec/op                    │           sec/op             vs base          │
BiogoIgor                                    13.53 ± 1%                    13.62 ± 1%  ~ (p=0.165 n=10)

          │ ./sweet/results/biogo-igor/baseline.results │ ./sweet/results/biogo-igor/experiment.results │
          │              average-RSS-bytes              │      average-RSS-bytes       vs base          │
BiogoIgor                                  62.19Mi ± 3%                  62.86Mi ± 1%  ~ (p=0.247 n=10)

          │ ./sweet/results/biogo-igor/baseline.results │ ./sweet/results/biogo-igor/experiment.results │
          │               peak-RSS-bytes                │       peak-RSS-bytes         vs base          │
BiogoIgor                                  89.57Mi ± 4%                  89.03Mi ± 3%  ~ (p=0.516 n=10)

          │ ./sweet/results/biogo-igor/baseline.results │ ./sweet/results/biogo-igor/experiment.results │
          │                peak-VM-bytes                │        peak-VM-bytes         vs base          │
BiogoIgor                                  766.4Mi ± 0%                  766.4Mi ± 0%  ~ (p=0.954 n=10)

shortname: sweet-biogo-krishna
             │ ./sweet/results/biogo-krishna/baseline.results │ ./sweet/results/biogo-krishna/experiment.results │
             │                     sec/op                     │          sec/op            vs base               │
BiogoKrishna                                       12.70 ± 2%                  12.09 ± 3%  -4.86% (p=0.000 n=10)

             │ ./sweet/results/biogo-krishna/baseline.results │ ./sweet/results/biogo-krishna/experiment.results │
             │               average-RSS-bytes                │       average-RSS-bytes         vs base          │
BiogoKrishna                                     4.085Gi ± 0%                     4.083Gi ± 0%  ~ (p=0.105 n=10)

             │ ./sweet/results/biogo-krishna/baseline.results │ ./sweet/results/biogo-krishna/experiment.results │
             │                 peak-RSS-bytes                 │         peak-RSS-bytes          vs base          │
BiogoKrishna                                     4.174Gi ± 0%                     4.173Gi ± 0%  ~ (p=0.853 n=10)

             │ ./sweet/results/biogo-krishna/baseline.results │ ./sweet/results/biogo-krishna/experiment.results │
             │                 peak-VM-bytes                  │         peak-VM-bytes           vs base          │
BiogoKrishna                                     4.877Gi ± 0%                     4.877Gi ± 0%  ~ (p=0.591 n=10)

shortname: sweet-bleve-index
                   │ ./sweet/results/bleve-index/baseline.results │ ./sweet/results/bleve-index/experiment.results │
                   │                    sec/op                    │            sec/op             vs base          │
BleveIndexBatch100                                     4.675 ± 1%                     4.669 ± 1%  ~ (p=0.739 n=10)

                   │ ./sweet/results/bleve-index/baseline.results │ ./sweet/results/bleve-index/experiment.results │
                   │              average-RSS-bytes               │      average-RSS-bytes        vs base          │
BleveIndexBatch100                                   185.5Mi ± 1%                   185.9Mi ± 1%  ~ (p=0.796 n=10)

                   │ ./sweet/results/bleve-index/baseline.results │ ./sweet/results/bleve-index/experiment.results │
                   │                peak-RSS-bytes                │        peak-RSS-bytes         vs base          │
BleveIndexBatch100                                   267.5Mi ± 6%                   265.0Mi ± 2%  ~ (p=0.739 n=10)

                   │ ./sweet/results/bleve-index/baseline.results │ ./sweet/results/bleve-index/experiment.results │
                   │                peak-VM-bytes                 │        peak-VM-bytes          vs base          │
BleveIndexBatch100                                   1.945Gi ± 4%                   1.945Gi ± 0%  ~ (p=0.725 n=10)

shortname: sweet-go-build
                    │ ./sweet/results/go-build/baseline.results │ ./sweet/results/go-build/experiment.results │
                    │                  sec/op                   │        sec/op         vs base               │
GoBuildKubelet                                       51.32 ± 0%             51.38 ± 3%       ~ (p=0.105 n=10)
GoBuildKubeletLink                                   7.669 ± 1%             7.663 ± 2%       ~ (p=0.579 n=10)
GoBuildIstioctl                                      46.02 ± 0%             46.07 ± 0%       ~ (p=0.739 n=10)
GoBuildIstioctlLink                                  8.174 ± 1%             8.143 ± 2%       ~ (p=0.436 n=10)
GoBuildFrontend                                      16.17 ± 1%             16.10 ± 1%       ~ (p=0.143 n=10)
GoBuildFrontendLink                                  1.399 ± 3%             1.377 ± 3%       ~ (p=0.218 n=10)
geomean                                              12.23                  12.18       -0.39%

shortname: sweet-gopher-lua
                     │ ./sweet/results/gopher-lua/baseline.results │ ./sweet/results/gopher-lua/experiment.results │
                     │                   sec/op                    │           sec/op             vs base          │
GopherLuaKNucleotide                                    22.71 ± 1%                    22.86 ± 1%  ~ (p=0.218 n=10)

                     │ ./sweet/results/gopher-lua/baseline.results │ ./sweet/results/gopher-lua/experiment.results │
                     │              average-RSS-bytes              │      average-RSS-bytes       vs base          │
GopherLuaKNucleotide                                  36.64Mi ± 2%                  36.40Mi ± 1%  ~ (p=0.631 n=10)

                     │ ./sweet/results/gopher-lua/baseline.results │ ./sweet/results/gopher-lua/experiment.results │
                     │               peak-RSS-bytes                │       peak-RSS-bytes         vs base          │
GopherLuaKNucleotide                                  43.28Mi ± 5%                  41.55Mi ± 7%  ~ (p=0.089 n=10)

                     │ ./sweet/results/gopher-lua/baseline.results │ ./sweet/results/gopher-lua/experiment.results │
                     │                peak-VM-bytes                │     peak-VM-bytes       vs base               │
GopherLuaKNucleotide                                  699.6Mi ± 0%             699.9Mi ± 0%  +0.04% (p=0.006 n=10)

shortname: sweet-markdown
                    │ ./sweet/results/markdown/baseline.results │ ./sweet/results/markdown/experiment.results │
                    │                  sec/op                   │          sec/op            vs base          │
MarkdownRenderXHTML                                 260.6m ± 4%                 256.4m ± 4%  ~ (p=0.796 n=10)

                    │ ./sweet/results/markdown/baseline.results │ ./sweet/results/markdown/experiment.results │
                    │             average-RSS-bytes             │     average-RSS-bytes      vs base          │
MarkdownRenderXHTML                                20.47Mi ± 1%                20.71Mi ± 2%  ~ (p=0.393 n=10)

                    │ ./sweet/results/markdown/baseline.results │ ./sweet/results/markdown/experiment.results │
                    │              peak-RSS-bytes               │      peak-RSS-bytes        vs base          │
MarkdownRenderXHTML                               20.88Mi ± 11%                21.73Mi ± 6%  ~ (p=0.470 n=10)

                    │ ./sweet/results/markdown/baseline.results │ ./sweet/results/markdown/experiment.results │
                    │               peak-VM-bytes               │       peak-VM-bytes        vs base          │
MarkdownRenderXHTML                                699.2Mi ± 0%                699.3Mi ± 0%  ~ (p=0.464 n=10)

shortname: sweet-tile38
                                   │ ./sweet/results/tile38/baseline.results │ ./sweet/results/tile38/experiment.results │
                                   │                 sec/op                  │       sec/op        vs base               │
Tile38WithinCircle100kmRequest                                   529.1µ ± 1%          530.3µ ± 1%       ~ (p=0.143 n=10)
Tile38IntersectsCircle100kmRequest                               629.6µ ± 1%          630.8µ ± 1%       ~ (p=0.971 n=10)
Tile38KNearestLimit100Request                                    446.4µ ± 1%          453.7µ ± 1%  +1.62% (p=0.000 n=10)
geomean                                                          529.8µ               533.4µ       +0.67%

                                   │ ./sweet/results/tile38/baseline.results │ ./sweet/results/tile38/experiment.results │
                                   │            average-RSS-bytes            │ average-RSS-bytes   vs base               │
Tile38WithinCircle100kmRequest                                  5.054Gi ± 1%         5.057Gi ± 1%       ~ (p=0.796 n=10)
Tile38IntersectsCircle100kmRequest                              5.381Gi ± 0%         5.431Gi ± 1%  +0.94% (p=0.019 n=10)
Tile38KNearestLimit100Request                                   6.801Gi ± 0%         6.802Gi ± 0%       ~ (p=0.684 n=10)
geomean                                                         5.697Gi              5.717Gi       +0.34%

                                   │ ./sweet/results/tile38/baseline.results │ ./sweet/results/tile38/experiment.results │
                                   │             peak-RSS-bytes              │   peak-RSS-bytes    vs base               │
Tile38WithinCircle100kmRequest                                  5.380Gi ± 1%         5.381Gi ± 1%       ~ (p=0.912 n=10)
Tile38IntersectsCircle100kmRequest                              5.669Gi ± 1%         5.756Gi ± 1%  +1.53% (p=0.019 n=10)
Tile38KNearestLimit100Request                                   7.013Gi ± 0%         7.011Gi ± 0%       ~ (p=0.796 n=10)
geomean                                                         5.980Gi              6.010Gi       +0.50%

                                   │ ./sweet/results/tile38/baseline.results │ ./sweet/results/tile38/experiment.results │
                                   │              peak-VM-bytes              │   peak-VM-bytes     vs base               │
Tile38WithinCircle100kmRequest                                  6.047Gi ± 1%         6.047Gi ± 1%       ~ (p=0.725 n=10)
Tile38IntersectsCircle100kmRequest                              6.305Gi ± 1%         6.402Gi ± 2%  +1.53% (p=0.035 n=10)
Tile38KNearestLimit100Request                                   7.685Gi ± 0%         7.685Gi ± 0%       ~ (p=0.955 n=10)
geomean                                                         6.642Gi              6.676Gi       +0.51%

                                   │ ./sweet/results/tile38/baseline.results │ ./sweet/results/tile38/experiment.results │
                                   │             p50-latency-sec             │  p50-latency-sec    vs base               │
Tile38WithinCircle100kmRequest                                   88.81µ ± 1%          89.36µ ± 1%  +0.61% (p=0.043 n=10)
Tile38IntersectsCircle100kmRequest                               151.5µ ± 1%          152.0µ ± 1%       ~ (p=0.089 n=10)
Tile38KNearestLimit100Request                                    259.0µ ± 0%          259.1µ ± 0%       ~ (p=0.853 n=10)
geomean                                                          151.6µ               152.1µ       +0.33%

                                   │ ./sweet/results/tile38/baseline.results │ ./sweet/results/tile38/experiment.results │
                                   │             p90-latency-sec             │  p90-latency-sec    vs base               │
Tile38WithinCircle100kmRequest                                   712.5µ ± 0%          713.9µ ± 1%       ~ (p=0.190 n=10)
Tile38IntersectsCircle100kmRequest                               960.6µ ± 1%          958.2µ ± 1%       ~ (p=0.739 n=10)
Tile38KNearestLimit100Request                                    1.007m ± 1%          1.032m ± 1%  +2.50% (p=0.000 n=10)
geomean                                                          883.4µ               890.5µ       +0.80%

                                   │ ./sweet/results/tile38/baseline.results │ ./sweet/results/tile38/experiment.results │
                                   │             p99-latency-sec             │  p99-latency-sec    vs base               │
Tile38WithinCircle100kmRequest                                   7.061m ± 1%          7.085m ± 1%       ~ (p=0.481 n=10)
Tile38IntersectsCircle100kmRequest                               7.228m ± 1%          7.187m ± 1%       ~ (p=0.143 n=10)
Tile38KNearestLimit100Request                                    2.085m ± 0%          2.131m ± 1%  +2.22% (p=0.000 n=10)
geomean                                                          4.738m               4.770m       +0.66%

                                   │ ./sweet/results/tile38/baseline.results │ ./sweet/results/tile38/experiment.results │
                                   │                  ops/s                  │       ops/s         vs base               │
Tile38WithinCircle100kmRequest                                   17.01k ± 1%          16.97k ± 1%       ~ (p=0.143 n=10)
Tile38IntersectsCircle100kmRequest                               14.29k ± 1%          14.27k ± 1%       ~ (p=0.988 n=10)
Tile38KNearestLimit100Request                                    20.16k ± 1%          19.84k ± 1%  -1.59% (p=0.000 n=10)
geomean                                                          16.99k               16.87k       -0.67%

shortname: uber_tally
goos: linux
goarch: amd64
pkg: github.com/uber-go/tally
cpu: Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz
                                │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                │                    sec/op                    │         sec/op           vs base               │
ScopeTaggedNoCachedSubscopes-12                                    2.867µ ± 4%               2.921µ ± 4%       ~ (p=0.579 n=10)
HistogramAllocation-12                                             1.519µ ± 3%               1.507µ ± 7%       ~ (p=0.631 n=10)
geomean                                                            2.087µ                    2.098µ       +0.53%

                       │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                       │                     B/op                     │             B/op              vs base          │
HistogramAllocation-12                                   1.124Ki ± 1%                   1.125Ki ± 4%  ~ (p=0.271 n=10)

                       │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                       │                  allocs/op                   │         allocs/op           vs base            │
HistogramAllocation-12                                     20.00 ± 0%                   20.00 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

shortname: uber_zap
pkg: go.uber.org/zap/zapcore
                                              │ ./bent-bench/20230303T173250.baseline.stdout │ ./bent-bench/20230303T173250.experiment.stdout │
                                              │                    sec/op                    │         sec/op          vs base                │
BufferedWriteSyncer/write_file_with_buffer-12                                   296.1n ± 12%             205.9n ± 10%  -30.46% (p=0.000 n=10)
MultiWriteSyncer/2_discarder-12                                                 7.528n ±  4%             7.014n ±  2%   -6.83% (p=0.000 n=10)
MultiWriteSyncer/4_discarder-12                                                 9.065n ±  1%             8.908n ±  1%   -1.73% (p=0.002 n=10)
MultiWriteSyncer/4_discarder_with_buffer-12                                     225.2n ±  2%             147.6n ±  2%  -34.48% (p=0.000 n=10)
WriteSyncer/write_file_with_no_buffer-12                                        4.785µ ±  1%             4.933µ ±  3%   +3.08% (p=0.001 n=10)
ZapConsole-12                                                                   702.5n ±  1%             649.1n ±  1%   -7.62% (p=0.000 n=10)
JSONLogMarshalerFunc-12                                                         1.219µ ±  2%             1.226µ ±  3%        ~ (p=0.781 n=10)
ZapJSON-12                                                                      555.4n ±  1%             480.9n ±  3%  -13.40% (p=0.000 n=10)
StandardJSON-12                                                                 814.1n ±  1%             809.0n ±  0%        ~ (p=0.101 n=10)
Sampler_Check/7_keys-12                                                         10.55n ±  2%             10.61n ±  1%        ~ (p=0.594 n=10)
Sampler_Check/50_keys-12                                                        11.01n ±  0%             10.98n ±  1%        ~ (p=0.286 n=10)
Sampler_Check/100_keys-12                                                       10.71n ±  0%             10.71n ±  0%        ~ (p=0.563 n=10)
Sampler_CheckWithHook/7_keys-12                                                 20.20n ±  2%             20.42n ±  2%        ~ (p=0.446 n=10)
Sampler_CheckWithHook/50_keys-12                                                20.72n ±  2%             21.02n ±  1%        ~ (p=0.078 n=10)
Sampler_CheckWithHook/100_keys-12                                               20.15n ±  2%             20.68n ±  3%   +2.63% (p=0.037 n=10)
TeeCheck-12                                                                     140.8n ±  2%             140.5n ±  2%        ~ (p=0.754 n=10)
geomean                                                                         87.80n                   82.39n         -6.15%

The only large regression (in ethereum_bitutil's BaseTest2KB) appears to
be spurious, as the test does not involve any goroutines (or
B.RunParallel()), which profiling confirms.

Updates golang/go#18237
Related to golang/go#32113
  • Loading branch information
@nixprime
nixprime committed Mar 3, 2023
1 parent cd6d225 commit 1bc1579
Show file tree
Hide file tree
Showing 5 changed files with 139 additions and 46 deletions.
2 changes: 1 addition & 1 deletion src/runtime/mgc.go
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1566,7 +1566,7 @@ func gcSweep(mode gcMode) {
lock(&sweep.lock)
if sweep.parked {
sweep.parked = false
ready(sweep.g, 0, true)
ready(sweep.g, 0, true, 0, true)
}
unlock(&sweep.lock)
}
Expand Down
2 changes: 1 addition & 1 deletion src/runtime/mgcmark.go
View file
Original file line number Diff line number Diff line change
Expand Up @@ -690,7 +690,7 @@ func gcFlushBgCredit(scanWork int64) {
// scheduler priority to get itself always run
// before other goroutines and always in the
// fresh quantum started by GC.
ready(gp, 0, false)
ready(gp, 0, false, 0, true)
} else {
// Partially satisfy this assist.
gp.gcAssistBytes += scanBytes
Expand Down
2 changes: 1 addition & 1 deletion src/runtime/preempt.go
View file
Original file line number Diff line number Diff line change
Expand Up @@ -274,7 +274,7 @@ func resumeG(state suspendGState) {

if state.stopped {
// We stopped it, so we need to re-schedule it.
ready(gp, 0, true)
ready(gp, 0, true, 0, true)
}
}

Expand Down
174 changes: 131 additions & 43 deletions src/runtime/proc.go
View file
Original file line number Diff line number Diff line change
Expand Up @@ -388,8 +388,18 @@ func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
}

func goready(gp *g, traceskip int) {
callergp := getg()
now := nanotime()
needWakep := readyUpdateAndPredictNeedWakep(callergp, now)
systemstack(func() {
ready(gp, traceskip, true, now, needWakep)
})
}

func goreadyNoWakep(gp *g, traceskip int) {
now := nanotime()
systemstack(func() {
ready(gp, traceskip, true)
ready(gp, traceskip, true, now, false /* needWakep */)
})
}

Expand Down Expand Up @@ -858,8 +868,42 @@ func fastrandinit() {
getRandomData(s)
}

func parkUpdateNeedWakep(gp *g) {
gp.firstReady = 0
if gp.lastReady != 0 {
gp.runnextSwitchHistory = nanotime()-gp.lastReady <= runnextSwitchNS
gp.lastReady = 0
}
}

// Returns true if gp, calling goready() or newproc() at now, is predicted
// *not* to park within the next runnextSwitchNS, such that a spinning P is
// needed to run the new G.
// Incorrectly returning true (causing a spinning P to be uselessly woken)
// wastes cycles but is harmless.
// Incorrectly returning false is handled by sysmon (in retake()).
func readyUpdateAndPredictNeedWakep(gp *g, now int64) bool {
pp := gp.m.p.ptr()

// Update predictor state.
gp.lastReady = now
if now-gp.firstReady > runnextSwitchNS {
if gp.firstReady != 0 {
gp.runnextSwitchHistory = false
}
gp.firstReady = now
}

// If the caller's runqueue is non-empty, predict that we need wakep();
// even if gp parks, there's no guarantee that following Gs will.
if !runqempty(pp) {
return true
}
return !gp.runnextSwitchHistory
}

// Mark gp ready to run.
func ready(gp *g, traceskip int, next bool) {
func ready(gp *g, traceskip int, next bool, now int64, needWakep bool) {
if trace.enabled {
traceGoUnpark(gp, traceskip)
}
Expand All @@ -875,8 +919,10 @@ func ready(gp *g, traceskip int, next bool) {

// status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
casgstatus(gp, _Gwaiting, _Grunnable)
runqput(mp.p.ptr(), gp, next)
wakep()
runqput(mp.p.ptr(), gp, next, now)
if needWakep {
wakep()
}
releasem(mp)
}

Expand Down Expand Up @@ -2715,7 +2761,7 @@ top:
// Wake up the finalizer G.
if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
if gp := wakefing(); gp != nil {
ready(gp, 0, true)
ready(gp, 0, true, now, true)
}
}
if *cgo_yield != nil {
Expand Down Expand Up @@ -3027,7 +3073,7 @@ func pollWork() bool {
func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
pp := getg().m.p.ptr()

ranTimer := false
ranTimerOrRunnextPending := false

const stealTries = 4
for i := 0; i < stealTries; i++ {
Expand Down Expand Up @@ -3072,16 +3118,21 @@ func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWo
// stolen G's. So check now if there
// is a local G to run.
if gp, inheritTime := runqget(pp); gp != nil {
return gp, inheritTime, now, pollUntil, ranTimer
return gp, inheritTime, now, pollUntil, ranTimerOrRunnextPending
}
ranTimer = true
ranTimerOrRunnextPending = true
}
}

// Don't bother to attempt to steal if p2 is idle.
if !idlepMask.read(enum.position()) {
if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
return gp, false, now, pollUntil, ranTimer
tnow, gp, runnextPending := runqsteal(pp, p2, stealTimersOrRunNextG, now)
now = tnow
if runnextPending {
ranTimerOrRunnextPending = true
}
if gp != nil {
return gp, false, now, pollUntil, ranTimerOrRunnextPending
}
}
}
Expand All @@ -3090,7 +3141,7 @@ func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWo
// No goroutines found to steal. Regardless, running a timer may have
// made some goroutine ready that we missed. Indicate the next timer to
// wait for.
return nil, false, now, pollUntil, ranTimer
return nil, false, now, pollUntil, ranTimerOrRunnextPending
}

// Check all Ps for a runnable G to steal.
Expand Down Expand Up @@ -3474,6 +3525,8 @@ func parkunlock_c(gp *g, lock unsafe.Pointer) bool {

// park continuation on g0.
func park_m(gp *g) {
parkUpdateNeedWakep(gp)

mp := getg().m

if trace.enabled {
Expand Down Expand Up @@ -3597,7 +3650,7 @@ func goyield_m(gp *g) {
pp := gp.m.p.ptr()
casgstatus(gp, _Grunning, _Grunnable)
dropg()
runqput(pp, gp, false)
runqput(pp, gp, false, 0)
schedule()
}

Expand Down Expand Up @@ -4232,11 +4285,13 @@ func newproc(fn *funcval) {
pc := getcallerpc()
systemstack(func() {
newg := newproc1(fn, gp, pc)
now := nanotime()
needWakep := readyUpdateAndPredictNeedWakep(gp, now)

pp := getg().m.p.ptr()
runqput(pp, newg, true)
runqput(pp, newg, true, now)

if mainStarted {
if mainStarted && needWakep {
wakep()
}
})
Expand Down Expand Up @@ -5389,6 +5444,7 @@ func sysmon() {
}
// retake P's blocked in syscalls
// and preempt long running G's
// and start P's if we mispredicted that wakep() was unnecessary
if retake(now) != 0 {
idle = 0
} else {
Expand Down Expand Up @@ -5424,6 +5480,7 @@ const forcePreemptNS = 10 * 1000 * 1000 // 10ms

func retake(now int64) uint32 {
n := 0
needWakep := false
// Prevent allp slice changes. This lock will be completely
// uncontended unless we're already stopping the world.
lock(&allpLock)
Expand All @@ -5437,6 +5494,9 @@ func retake(now int64) uint32 {
// allp but not yet created new Ps.
continue
}
if runqstealable(pp, now) && sched.nmspinning.Load() == 0 {
needWakep = true
}
pd := &pp.sysmontick
s := pp.status
sysretake := false
Expand Down Expand Up @@ -5488,6 +5548,10 @@ func retake(now int64) uint32 {
}
}
unlock(&allpLock)
if needWakep {
wakep()
n++
}
return uint32(n)
}

Expand Down Expand Up @@ -5766,7 +5830,7 @@ func globrunqget(pp *p, max int32) *g {
n--
for ; n > 0; n-- {
gp1 := sched.runq.pop()
runqput(pp, gp1, false)
runqput(pp, gp1, false, 0)
}
return gp
}
Expand Down Expand Up @@ -5933,6 +5997,25 @@ func runqempty(pp *p) bool {
}
}

// runqstealable is like !runqempty, but returns false if pp has a G in
// p.runnext that can't be stolen yet.
func runqstealable(pp *p, now int64) bool {
for {
head := atomic.Load(&pp.runqhead)
tail := atomic.Load(&pp.runqtail)
runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
if tail == atomic.Load(&pp.runqtail) {
if head != tail {
return true
}
if runnext == 0 {
return false
}
return ((guintptr)(runnext)).ptr().runnextSince+runnextSwitchNS < now
}
}
}

// To shake out latent assumptions about scheduling order,
// we introduce some randomness into scheduling decisions
// when running with the race detector.
Expand All @@ -5949,12 +6032,16 @@ const randomizeScheduler = raceenabled
// If next is true, runqput puts g in the pp.runnext slot.
// If the run queue is full, runnext puts g on the global queue.
// Executed only by the owner P.
func runqput(pp *p, gp *g, next bool) {
func runqput(pp *p, gp *g, next bool, now int64) {
if randomizeScheduler && next && fastrandn(2) == 0 {
next = false
}

if next {
if now == 0 {
now = nanotime()
}
gp.runnextSince = now
retryNext:
oldnext := pp.runnext
if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
Expand Down Expand Up @@ -6123,11 +6210,22 @@ retry:
return
}

// Don't steal p.runnext if it's been made runnable within the last
// runnextSwitchNS and the P is running.
// The important use case here is when the G running on the P ready()s another G
// and then almost immediately blocks.
// Giving the P a chance to schedule runnext avoids thrashing Gs between
// different Ps.
// On most platforms, sleep timeout granularity is coarser than
// runnextSwitchNS, so sleeping will result in significant overshoot; instead,
// stealWork() instructs findRunnable() to spin-wait.
const runnextSwitchNS = 5e3

// Grabs a batch of goroutines from pp's runnable queue into batch.
// Batch is a ring buffer starting at batchHead.
// Returns number of grabbed goroutines.
// Can be executed by any P.
func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool, now int64) (rnow int64, n uint32, runnextPending bool) {
for {
h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
Expand All @@ -6138,33 +6236,22 @@ func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool)
// Try to steal from pp.runnext.
if next := pp.runnext; next != 0 {
if pp.status == _Prunning {
// Sleep to ensure that pp isn't about to run the g
// we are about to steal.
// The important use case here is when the g running
// on pp ready()s another g and then almost
// immediately blocks. Instead of stealing runnext
// in this window, back off to give pp a chance to
// schedule runnext. This will avoid thrashing gs
// between different Ps.
// A sync chan send/recv takes ~50ns as of time of
// writing, so 3us gives ~50x overshoot.
if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
usleep(3)
} else {
// On some platforms system timer granularity is
// 1-15ms, which is way too much for this
// optimization. So just yield.
osyield()
// Enforce runnextSwitchNS.
if now == 0 {
now = nanotime()
}
if now <= next.ptr().runnextSince+runnextSwitchNS {
return now, 0, true
}
}
if !pp.runnext.cas(next, 0) {
continue
}
batch[batchHead%uint32(len(batch))] = next
return 1
return now, 1, false
}
}
return 0
return now, 0, false
}
if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
continue
Expand All @@ -6174,31 +6261,32 @@ func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool)
batch[(batchHead+i)%uint32(len(batch))] = g
}
if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
return n
return now, n, false
}
}
}

// Steal half of elements from local runnable queue of p2
// and put onto local runnable queue of p.
// Returns one of the stolen elements (or nil if failed).
func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
func runqsteal(pp, p2 *p, stealRunNextG bool, now int64) (rnow int64, gp *g, runnextPending bool) {
t := pp.runqtail
n := runqgrab(p2, &pp.runq, t, stealRunNextG)
tnow, n, runnextPending := runqgrab(p2, &pp.runq, t, stealRunNextG, now)
now = tnow
if n == 0 {
return nil
return now, nil, runnextPending
}
n--
gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
gp = pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
if n == 0 {
return gp
return now, gp, runnextPending
}
h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
if t-h+n >= uint32(len(pp.runq)) {
throw("runqsteal: runq overflow")
}
atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
return gp
return now, gp, runnextPending
}

// A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
Expand Down
Loading

0 comments on commit 1bc1579

Please sign in to comment.