On Monday, May 31, 2021 at 10:07:35 AM UTC-4 axel.wa...@googlemail.com
wrote:
> On Mon, May 31, 2021 at 3:30 PM tapi...@gmail.com <tapi...@gmail.com>
> wrote:
>
>> On Sunday, May 30, 2021 at 12:54:02 PM UTC-4 axel.wa...@googlemail.com
>> wrote:
>>
>>> That is very normal for micro-benchmarks on a ns scale.
>>>
>> The results are so constantly that I think it is more related to CPU
>> cache and the specified directives.
>>
>
> Yes that's what I meant. Given that the signal is so small (on a ns
> scale), noise introduced by alignment and code caches and other
> micro-architectural details becomes so large (relatively), that it's normal
> to notice an effect in reordering or removing unrelated code.
>
> That's why they are not a good basis to base decisions on. You can't know
> if what you're seeing is a real effect or just noise.
>
I tend to think the result is reliable, though I agree they are not a good
basis to base decisions on,
for the performance might be different between different CPU architectures.
And it is still interesting to see the concrete CPU instructions generated
for different situations.
For example, if the source values are moved out of the benchmark functions
as package-level variables,
the generated instructions will become to
0x0034 00052 (/home/d630/Desktop/aa/t/x_test.go:11) MOVQ
"".struct3_1(SB), DX
0x003b 00059 (/home/d630/Desktop/aa/t/x_test.go:11) MOVQ
"".struct3_1+8(SB), BX
0x0042 00066 (/home/d630/Desktop/aa/t/x_test.go:11) MOVQ
"".struct3_1+16(SB), SI
0x0049 00073 (/home/d630/Desktop/aa/t/x_test.go:11) MOVQ DX,
"".struct3_0(SB)
0x0050 00080 (/home/d630/Desktop/aa/t/x_test.go:11) MOVQ BX,
"".struct3_0+8(SB)
0x0057 00087 (/home/d630/Desktop/aa/t/x_test.go:11) MOVQ SI,
"".struct3_0+16(SB)
and
0x0034 00052 (/home/d630/Desktop/aa/t/x_test.go:21) MOVQ
"".struct4_1(SB), DX
0x003b 00059 (/home/d630/Desktop/aa/t/x_test.go:21) MOVQ
"".struct4_1+8(SB), BX
0x0042 00066 (/home/d630/Desktop/aa/t/x_test.go:21) MOVQ
"".struct4_1+16(SB), SI
0x0049 00073 (/home/d630/Desktop/aa/t/x_test.go:21) MOVQ
"".struct4_1+24(SB), DI
0x0050 00080 (/home/d630/Desktop/aa/t/x_test.go:21) MOVQ DX,
"".struct4_0(SB)
0x0057 00087 (/home/d630/Desktop/aa/t/x_test.go:21) MOVQ BX,
"".struct4_0+8(SB)
0x005e 00094 (/home/d630/Desktop/aa/t/x_test.go:21) MOVQ SI,
"".struct4_0+16(SB)
0x0065 00101 (/home/d630/Desktop/aa/t/x_test.go:21) MOVQ DI,
"".struct4_0+24(SB)
And the benchmark result is more predictable:
Benchmark_CopyStruct_3_fields-4 988590592 1.441 ns/op
Benchmark_CopyStruct_4_fields-4 597097029 2.590 ns/op
>
>
>>
>>
>>>
>>> On Sun, May 30, 2021 at 6:07 PM tapi...@gmail.com <tapi...@gmail.com>
>>> wrote:
>>>
>>>> It is some strange that if any of the bool/byte/int16/int64 benchmarks
>>>> is removed in this test file https://play.golang.org/p/w29J9VhtzYH,
>>>> then the benchmark result is like:
>>>>
>>>> Benchmark_CopyStruct_3_fields-4 1000000000 0.7780
>>>> ns/op
>>>> Benchmark_CopyStruct_4_fields-4 1000000000 0.7513
>>>> ns/op
>>>>
>>>> Otherwise, it is like:
>>>>
>>>> Benchmark_CopyStruct_3_fields-4 1000000000 1.501 ns/op
>>>> Benchmark_CopyStruct_4_fields-4 1000000000 0.7755
>>>> ns/op
>>>>
>>>> In other words, it looks there is mutual interference between
>>>> benchmarks.
>>>>
>>>> On Saturday, May 29, 2021 at 11:49:47 PM UTC-4 tapi...@gmail.com wrote:
>>>>
>>>>> The benchmark code: https://play.golang.org/p/bC1zO14eNeh
>>>>>
>>>>> The result:
>>>>>
>>>>> $ go test -bench=.
>>>>> goos: linux
>>>>> goarch: amd64
>>>>> pkg: example.com/valuecopy
>>>>> cpu: Intel(R) Core(TM) i5-3210M CPU @ 2.50GHz
>>>>> Benchmark_CopyBool-4 1000000000 0.8885
>>>>> ns/op
>>>>> Benchmark_CopyByte-4 1000000000 0.8872
>>>>> ns/op
>>>>> Benchmark_CopyInt16-4 1000000000 0.8785
>>>>> ns/op
>>>>> Benchmark_CopyInt32-4 1000000000 0.8854
>>>>> ns/op
>>>>> Benchmark_CopyInt64-4 1000000000 0.8831
>>>>> ns/op
>>>>> Benchmark_CopyPointer-4 911733464 1.330
>>>>> ns/op
>>>>> Benchmark_CopyString-4 901249356 1.325
>>>>> ns/op
>>>>> Benchmark_CopySlice-4 664187247 1.765
>>>>> ns/op
>>>>> Benchmark_CopyArray_2_elements-4 1000000000 0.8874
>>>>> ns/op
>>>>> Benchmark_CopyArray_3_elements-4 1000000000 1.096
>>>>> ns/op
>>>>> Benchmark_CopyArray_4_elements-4 1000000000 1.105
>>>>> ns/op
>>>>> Benchmark_CopyArray_5_elements-4 534542524 2.202
>>>>> ns/op
>>>>> Benchmark_CopyArray_6_elements-4 727849554 1.606
>>>>> ns/op
>>>>> Benchmark_CopyArray_7_elements-4 444494692 2.649
>>>>> ns/op
>>>>> Benchmark_CopyArray_8_elements-4 584854867 1.993
>>>>> ns/op
>>>>> Benchmark_CopyArray_9_elements-4 389639859 3.083
>>>>> ns/op
>>>>> Benchmark_CopyArray_10_elements-4 267380602 4.418
>>>>> ns/op
>>>>> Benchmark_CopyArray_11_elements-4 242644033 4.867
>>>>> ns/op
>>>>> Benchmark_CopyArray_12_elements-4 268304104 4.498
>>>>> ns/op
>>>>> Benchmark_CopyArray_13_elements-4 82165272 14.46 ns/op
>>>>> Benchmark_CopyStruct_2_fields-4 1000000000 0.5029
>>>>> ns/op
>>>>> Benchmark_CopyStruct_3_fields-4 671136589 1.769
>>>>> ns/op
>>>>> Benchmark_CopyStruct_4_fields-4 1000000000 0.8785
>>>>> ns/op
>>>>> Benchmark_CopyStruct_5_fields-4 530876049 2.202
>>>>> ns/op
>>>>> Benchmark_CopyStruct_6_fields-4 723380257 1.581
>>>>> ns/op
>>>>> Benchmark_CopyStruct_7_fields-4 444619906 2.636
>>>>> ns/op
>>>>> Benchmark_CopyStruct_8_fields-4 588605260 1.968
>>>>> ns/op
>>>>> Benchmark_CopyStruct_9_fields-4 387253551 3.073
>>>>> ns/op
>>>>> Benchmark_CopyStruct_10_fields-4 267450452 4.396
>>>>> ns/op
>>>>> Benchmark_CopyStruct_11_fields-4 246289522 4.855
>>>>> ns/op
>>>>> Benchmark_CopyStruct_12_fields-4 266212528 4.426
>>>>> ns/op
>>>>> Benchmark_CopyStruct_13_fields-4 207298701 5.739
>>>>> ns/op
>>>>>
>>>>> From the benchmark result, it looks
>>>>> * the cost of copying a [13]int value is much larger than copying a
>>>>> [12]int value.
>>>>> * the cost of copying a struct{a, b, c int} value is about double of
>>>>> copying a struct{a, b, c, d int} value.
>>>>>
>>>>>
>>>>> --
>>>> You received this message because you are subscribed to the Google
>>>> Groups "golang-nuts" group.
>>>> To unsubscribe from this group and stop receiving emails from it, send
>>>> an email to golang-nuts...@googlegroups.com.
>>>>
>>> To view this discussion on the web visit
>>>> https://groups.google.com/d/msgid/golang-nuts/f53a8f05-37a6-4c86-87c1-de9e83f381fdn%40googlegroups.com
>>>>
>>>> <https://groups.google.com/d/msgid/golang-nuts/f53a8f05-37a6-4c86-87c1-de9e83f381fdn%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>> .
>>>>
>>> --
>> You received this message because you are subscribed to the Google Groups
>> "golang-nuts" group.
>> To unsubscribe from this group and stop receiving emails from it, send an
>> email to golang-nuts...@googlegroups.com.
>>
> To view this discussion on the web visit
>> https://groups.google.com/d/msgid/golang-nuts/6caafbc0-66f3-4b77-ac92-6174b17ebf30n%40googlegroups.com
>>
>> <https://groups.google.com/d/msgid/golang-nuts/6caafbc0-66f3-4b77-ac92-6174b17ebf30n%40googlegroups.com?utm_medium=email&utm_source=footer>
>> .
>>
>
--
You received this message because you are subscribed to the Google Groups
"golang-nuts" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to golang-nuts+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/golang-nuts/62f3863c-5d6c-4dc5-9d7a-47175217cf61n%40googlegroups.com.
