runtime: request for comment: reducing preemption in suspendG when G is running large nosplit functions

[shawndx]

What version of Go are you using (go version)?

$ go version
tip version, on arm64

Does this issue reproduce with the latest release?

Yes.

What operating system and processor architecture are you using (go env)?

go env Output

$ go env

GO111MODULE=""
GOARCH="arm64"
GOBIN=""
GOCACHE="/home/xiaji01/.cache/go-build"
GOENV="/home/xiaji01/.config/go/env"
GOEXE=""
GOFLAGS=""
GOHOSTARCH="arm64"
GOHOSTOS="linux"
GOINSECURE=""
GOMODCACHE="/home/xiaji01/.go/pkg/mod"
GONOPROXY=""
GONOSUMDB=""
GOOS="linux"
GOPATH="/home/xiaji01/.go"
GOPRIVATE=""
GOPROXY="https://proxy.golang.org,direct"
GOROOT="/home/xiaji01/src/go.gc"
GOSUMDB="sum.golang.org"
GOTMPDIR=""
GOTOOLDIR="/home/xiaji01/src/go.gc/pkg/tool/linux_arm64"
GCCGO="gccgo"
AR="ar"
CC="gcc"
CXX="g++"
CGO_ENABLED="1"
GOMOD="/home/xiaji01/src/go.gc/src/go.mod"
CGO_CFLAGS="-g -O2"
CGO_CPPFLAGS=""
CGO_CXXFLAGS="-g -O2"
CGO_FFLAGS="-g -O2"
CGO_LDFLAGS="-g -O2"
PKG_CONFIG="pkg-config"
GOGCCFLAGS="-fPIC -pthread -fmessage-length=0 -fdebug-prefix-map=/tmp/go-build944660995=/tmp/go-build -gno-record-gcc-switches"

What did you do?

live-lock problems are spotted on arm64 in a couple of micro benchmarks from text/tabwriter (there should be more similar cases).
If a goroutine is running 'nosplit' functions which are time consuming, like
bulkBarrierPreWriteSrcOnly
memmove
called in growslice when the slice is a large one, and its background worker tries to suspend it to do stack scanning, the G under preemption may hardly make progress, especially on arm64 machines which seem to experience poor signal handling performance.

pprof data of BenchmarkTable/1x100000/new on an arm64 machine:

(pprof) top
Showing nodes accounting for 12.46s, 95.70% of 13.02s total
Dropped 118 nodes (cum <= 0.07s)
Showing top 10 nodes out of 40
      flat  flat%   sum%        cum   cum%
     5.41s 41.55% 41.55%      5.41s 41.55%  runtime.futex
     3.22s 24.73% 66.28%      3.59s 27.57%  runtime.nanotime (inline)
     0.99s  7.60% 73.89%      0.99s  7.60%  runtime.tgkill
     0.84s  6.45% 80.34%      0.84s  6.45%  runtime.osyield
     0.77s  5.91% 86.25%      0.77s  5.91%  runtime.epollwait
     0.37s  2.84% 89.09%      0.37s  2.84%  runtime.nanotime1
     0.33s  2.53% 91.63%      6.38s 49.00%  runtime.suspendG
     0.30s  2.30% 93.93%      0.30s  2.30%  runtime.getpid
     0.14s  1.08% 95.01%      0.14s  1.08%  runtime.procyield
     0.09s  0.69% 95.70%      1.38s 10.60%  runtime.preemptM (inline)

I'm thinking to introduce a flag into G to indicate whether it's running a time-consuming nosplit function and let suspendG yield to avoid the live-lock, the flag is set manually for known functions, I tried it for bulkBarrierPreWriteSrcOnly and memmove in runtime.growslice and runtime.makeslicecopy when the slice is longer than a threshold (select 4K for now), the text/tabwriter package witnessed significant improvement on arm64:

arm64-1 perf:

name                        old time/op    new time/op    delta
Table/1x10/new-224            6.59µs ± 8%    6.67µs ± 7%     ~     (p=0.841 n=5+5)
Table/1x10/reuse-224          1.94µs ± 1%    1.94µs ± 1%     ~     (p=0.802 n=5+5)
Table/1x1000/new-224           494µs ±13%     381µs ± 3%  -22.85%  (p=0.008 n=5+5)
Table/1x1000/reuse-224         185µs ± 0%     185µs ± 0%   -0.37%  (p=0.016 n=5+4)
Table/1x100000/new-224         2.73s ±73%     0.04s ± 2%  -98.46%  (p=0.008 n=5+5)
Table/1x100000/reuse-224       4.25s ±76%     0.02s ± 1%  -99.55%  (p=0.008 n=5+5)
Table/10x10/new-224           20.7µs ± 6%    20.1µs ± 5%     ~     (p=0.421 n=5+5)
Table/10x10/reuse-224         8.89µs ± 0%    8.91µs ± 0%   +0.21%  (p=0.032 n=5+5)
Table/10x1000/new-224         1.70ms ± 7%    1.59ms ± 6%     ~     (p=0.151 n=5+5)
Table/10x1000/reuse-224        908µs ± 0%     902µs ± 0%   -0.59%  (p=0.032 n=5+5)
Table/10x100000/new-224        2.34s ±62%     0.14s ± 2%  -94.21%  (p=0.008 n=5+5)
Table/10x100000/reuse-224      911ms ±72%     103ms ± 2%  -88.66%  (p=0.008 n=5+5)
Table/100x10/new-224           143µs ±15%     138µs ± 5%     ~     (p=0.841 n=5+5)
Table/100x10/reuse-224        78.1µs ± 0%    78.1µs ± 0%     ~     (p=0.841 n=5+5)
Table/100x1000/new-224        13.1ms ± 6%    11.9ms ± 5%   -9.16%  (p=0.008 n=5+5)
Table/100x1000/reuse-224      8.13ms ± 0%    8.15ms ± 1%     ~     (p=1.000 n=5+5)
Table/100x100000/new-224       1.29s ±17%     1.31s ±10%     ~     (p=0.310 n=5+5)
Table/100x100000/reuse-224     1.23s ± 2%     1.26s ± 5%     ~     (p=0.286 n=4+5)

arm64-2 perf:

name                       old time/op    new time/op    delta
Table/1x10/new-64            4.33µs ± 1%    5.40µs ± 4%  +24.92%  (p=0.004 n=6+5)
Table/1x10/reuse-64          1.77µs ± 0%    1.77µs ± 0%   -0.36%  (p=0.024 n=6+6)
Table/1x1000/new-64           318µs ±10%     379µs ±10%  +19.04%  (p=0.008 n=5+5)
Table/1x1000/reuse-64         172µs ± 0%     171µs ± 0%     ~     (p=0.537 n=6+5)
Table/1x100000/new-64         4.26s ±88%     0.05s ± 4%  -98.77%  (p=0.002 n=6+6)
Table/1x100000/reuse-64       4.79s ±73%     0.02s ± 2%  -99.61%  (p=0.004 n=6+5)
Table/10x10/new-64           14.6µs ± 3%    16.4µs ± 9%  +12.51%  (p=0.004 n=5+6)
Table/10x10/reuse-64         8.76µs ± 0%    8.79µs ± 0%     ~     (p=0.329 n=5+6)
Table/10x1000/new-64         1.18ms ± 3%    1.32ms ± 4%  +12.12%  (p=0.002 n=6+6)
Table/10x1000/reuse-64        890µs ± 0%     897µs ± 0%   +0.84%  (p=0.002 n=6+6)
Table/10x100000/new-64        1.43s ±36%     0.16s ± 3%  -88.56%  (p=0.004 n=5+6)
Table/10x100000/reuse-64      375ms ±55%     138ms ±11%  -63.31%  (p=0.004 n=5+6)
Table/100x10/new-64           103µs ± 0%     105µs ± 4%     ~     (p=0.429 n=5+6)
Table/100x10/reuse-64        79.6µs ± 0%    80.0µs ± 1%     ~     (p=0.329 n=5+6)
Table/100x1000/new-64        10.2ms ± 3%    10.5ms ± 1%     ~     (p=0.052 n=6+5)
Table/100x1000/reuse-64      8.75ms ± 0%    9.23ms ± 6%     ~     (p=0.126 n=5+6)
Table/100x100000/new-64       1.42s ± 6%     1.46s ±11%     ~     (p=0.394 n=6+6)
Table/100x100000/reuse-64     1.40s ± 0%     1.45s ± 6%     ~     (p=0.690 n=5+5)
Pyramid/10-64                15.3µs ±45%    14.5µs ±20%     ~     (p=0.662 n=6+5)
Pyramid/100-64               1.16ms ±13%    0.88ms ±11%  -23.90%  (p=0.002 n=6+6)
Pyramid/1000-64              71.6ms ± 8%    77.8ms ±12%     ~     (p=0.056 n=5+5)
Ragged/10-64                 13.2µs ±17%    14.0µs ± 2%     ~     (p=0.792 n=6+5)
Ragged/100-64                 104µs ± 3%     123µs ± 2%  +18.21%  (p=0.008 n=5+5)
Ragged/1000-64               1.12ms ±12%    1.35ms ± 9%  +20.55%  (p=0.002 n=6+6)
Code-64                      3.51µs ± 1%    3.80µs ± 3%   +8.46%  (p=0.004 n=5+6)

x86 perf:

name                       old time/op    new time/op    delta
Table/1x10/new-32            5.36µs ± 2%    5.58µs ± 3%   +4.03%  (p=0.004 n=6+6)
Table/1x10/reuse-32          1.50µs ± 8%    1.48µs ± 0%     ~     (p=0.433 n=6+5)
Table/1x1000/new-32           318µs ± 2%     355µs ± 2%  +11.43%  (p=0.004 n=5+6)
Table/1x1000/reuse-32         127µs ± 0%     134µs ± 1%   +5.36%  (p=0.004 n=6+5)
Table/1x100000/new-32        49.9ms ± 3%    48.1ms ± 3%   -3.57%  (p=0.026 n=6+6)
Table/1x100000/reuse-32      13.8ms ± 1%    15.2ms ± 1%   +9.87%  (p=0.002 n=6+6)
Table/10x10/new-32           17.4µs ± 3%    17.6µs ± 2%     ~     (p=0.310 n=6+6)
Table/10x10/reuse-32         7.20µs ± 1%    7.21µs ± 0%     ~     (p=0.429 n=5+6)
Table/10x1000/new-32         1.38ms ± 1%    1.43ms ± 2%   +4.16%  (p=0.002 n=6+6)
Table/10x1000/reuse-32        687µs ± 1%     693µs ± 0%     ~     (p=0.052 n=6+5)
Table/10x100000/new-32        131ms ± 2%     133ms ± 5%     ~     (p=0.699 n=6+6)
Table/10x100000/reuse-32     89.2ms ± 2%    90.3ms ± 2%     ~     (p=0.177 n=5+6)
Table/100x10/new-32           122µs ± 1%     122µs ± 1%     ~     (p=0.632 n=6+5)
Table/100x10/reuse-32        62.3µs ± 0%    62.4µs ± 0%     ~     (p=0.429 n=5+6)
Table/100x1000/new-32        12.2ms ± 4%    12.0ms ± 3%     ~     (p=0.180 n=6+6)
Table/100x1000/reuse-32      6.29ms ± 0%    6.32ms ± 0%   +0.54%  (p=0.015 n=6+6)
Table/100x100000/new-32       1.01s ± 2%     1.00s ± 4%     ~     (p=0.429 n=6+5)
Table/100x100000/reuse-32     972ms ±10%     962ms ±18%     ~     (p=1.000 n=6+6)
Pyramid/10-32                14.3µs ± 4%    14.8µs ± 6%     ~     (p=0.240 n=6+6)
Pyramid/100-32                833µs ± 2%     841µs ± 2%     ~     (p=0.394 n=6+6)
Pyramid/1000-32              53.5ms ± 1%    56.2ms ± 3%   +4.98%  (p=0.004 n=5+6)
Ragged/10-32                 15.1µs ± 0%    14.8µs ± 1%   -1.75%  (p=0.008 n=5+5)
Ragged/100-32                 127µs ± 4%     130µs ± 2%     ~     (p=0.180 n=6+6)
Ragged/1000-32               1.27ms ± 5%    1.33ms ± 3%   +3.96%  (p=0.026 n=6+6)
Code-32                      3.75µs ± 1%    3.81µs ± 2%   +1.67%  (p=0.048 n=5+6)

The potential overhead is acquiring the current G in the two slice functions, which seem to be tiny.
I'm working on benchmarking more packages and evaluate its impact to x86 platform.

Any comment is highly appreciated.

What did you expect to see?

Better performance.

What did you see instead?

[ianlancetaylor]

CC @aclements @cherrymui

[cherrymui]

I think the plan is the opposite: make functions like memmove preemptible. See also #36365

[shawndx]

I think the plan is the opposite: make functions like memmove preemptible. See also #36365

Thanks @cherrymui , do we have a (tentative) scheduled version to have it?