善用 Go 語言效能測試工具來提升執行效率

在 AI 訓練模型前，都需要經過大量的資料處理，而資料處理的速度在整個流程內扮演很重要的角色，寫出高效能的 Parser 能降低整體處理時間，那如何評估程式效能如何，以及如何快速找到效能瓶頸？本議程會帶大家了解 Go 語言內建的效能測試工具，透過 Benchmark 來找出程式效能瓶頸的地方，快速改善及優化，讓整個系統流程更順暢。也會順道分享 Go 在字串處理優化的一些小技巧。聽過此議程相信您對 Go 語言會有更深入的了解，如果你想寫出有效率的程式碼，本議程一定不能錯過。

Transcript

1. 善⽤ Go 語⾔效能測試
   ⼯具來提升執⾏效率
   Bo-Yi Wu
   2020/10/24
   

   View Slide

2. About me
   • Software Engineer in Mediatek
   • Member of Drone CI/CD Platform
   • Member of Gitea Platform
   • Member of Gin Golang Framework
   • Maintain Some GitHub Actions Plugins.
   

   View Slide

3. 使⽤ AI 來偵測異常
   減少晶片設計發⽣的問題
   

   View Slide

4. View Slide

5. Data 資料量
   • 1 ~ 2 萬個檔案
   • 每個檔案 1000 萬⾏以上
   • 總共 8 億 ~ 1x 億個資料點 (單⼀製程)
   

   View Slide

6. 效能結果 (原先使⽤ 400 台機器)
   1BSTFS ,# .# .
   (# .)8
   
   PME
   
   
   
   
   
   VOJUTFDPOE
   ఏঋYYഒ
   DPSF(
   লԼ)8㑌࣍୆(3BN7.
   

   View Slide

7. Python to Golang
   ఏঋ଎౓۰౸ළᰍ
   

   View Slide

8. 效能結果 (原先使⽤ 400 台機器)
   1BSTFS ,# .# .
   (# .)8
   
   PME
   
   
   
   
   
   VOJUTFDPOE
   ఏঋYYഒ
   DPSF(
   লԼ)8㑌࣍୆(3BN7.
   

   View Slide

9. Write Parser in Golang
   ᅷยိݯIUUQT[IVBOMBO[IJIVDPNQ
   

   View Slide

10. %BUB㚎༰֨ࣜ
    

    View Slide

11. ⾃⾏研發 Parser 原因
    • 提⾼部⾨產能 (太多單位有分析需求)
    • 減少 Computing Farm 使⽤ (省錢)
    • 整合 AI Model 花費時間過久 (inference time
    不到⼀秒啊)
    • ⽬前在市⾯上找不到任何⼀套開源 Parser?
    

    View Slide

12. 效能結果 (原先使⽤ 400 台機器)
    1BSTFS ,# .# .
    (# .)8
    
    PME
    
    
    
    
    
    VOJUTFDPOE
    ఏঋYYഒ
    DPSF(
    লԼ)8㑌࣍୆(3BN7.
    

    View Slide

13. 第⼀個版本
    var group = regexp.MustCompile(`^(\s*)(\S+)\s*\((.*?)\)\s*{\s*$`)
    var groupDone = regexp.MustCompile(`^\s*}\s*$`)
    /// Simple attribute compile.
    /// key : value;
    var simpleAttribute = regexp.MustCompile(`^(\s*)(\S+)\s*:\s*(.+)\s*;.*$`)
    var specialSimpleAttribute = regexp.MustCompile(`^(\s*)(\S+)\s*:\s*(.+)\s*$`)
    /// Complex attribute compile.
    /// key (valueList);
    var complexAttribute = regexp.MustCompile(`^(\s*)(\S+)\s*(\(.+\))\s*;.*$`)
    var specialComplexAttribute = regexp.MustCompile(`^(\s*)(\S+)\s*(\(.+\))\s*$`)
    Ꮘೳළᰍ
    

    View Slide

14. 效能評估⽅式
    

    View Slide

15. 記錄程式執⾏時間
    $ time your_program
    SFBMNT
    VTFSNT
    TZTNT
    

    View Slide

16. 埋入程式碼
    func main() {
    start := time.Now()
    fmt.Println(foobar(1000, 10))
    elapsed := time.Since(start)
    log.Printf("foobar took %s", elapsed)
    }
    

    View Slide

17. Go 內建性能⼯具
    

    View Slide

18. Go benchmark
    $ go test -v -bench=foobar -run=none
    

    View Slide

19. var fixtures = []struct {
    path string
    }{
    {"cells.lib"},
    {"cells_timing.lib"},
    {"sample2_Late.lib"},
    {"small.lib"},
    }
    func withFixtures(b *testing.B, fn func(b *testing.B, input []byte)) {
    for _, fix ::= range fixtures {
    data, err ::= ioutil.ReadFile(filepath.Join(".../testdata", fix.path))
    if err !!= nil {
    b.Fatal(err)
    }
    b.Run(fix.path, func(b *testing.B) {
    b.ReportAllocs()
    b.SetBytes(int64(len(data)))
    b.ResetTimer()
    fn(b, data)
    })
    }
    }
    3VO1BSBMMFMฏߦ႔ཧ
    3FQPSU"MMPDTهԱᱪ࢖༻ྔ
    4FU#ZUFTᄸ࣍॥؀࢖༻ଟগهԱᱪ
    3FTFU5JNFSॏஔ5JNFSܭࢉ
    

    View Slide

20. func BenchmarkWhitespaceArrayInlined(b *testing.B) {
    withFixtures(b, benchmarkWhitespaceArrayInlined)
    }
    func BenchmarkWhitespaceArrayLoop(b *testing.B) {
    withFixtures(b, benchmarkWhitespaceArray)
    }
    func BenchmarkWhitespaceSwitch(b *testing.B) {
    withFixtures(b, benchmarkWhitespaceSwitch)
    }
    func BenchmarkWhitespaceIf(b *testing.B) {
    withFixtures(b, benchmarkWhitespaceIf)
    }
    func BenchmarkWhitespaceIfInlined(b *testing.B) {
    withFixtures(b, benchmarkWhitespaceIfInlined)
    }
    ଟ䈕Ҋෆಉํ๏Ꮘೳଌࢼ
    

    View Slide

21. View Slide

22. RunParallel (平⾏處理)
    根據系統 CPU 數量加速處理
    HPUFTUDQV SVO?CFODICFODINBSL@UFTUHP
    

    View Slide

23. n ::= uint64(0)
    numProcs ::= b.parallelism * runtime.GOMAXPROCS(0)
    var wg sync.WaitGroup
    wg.Add(numProcs)
    for p ::= 0; p < numProcs; p+++ {
    go func() {
    defer wg.Done()
    pb ::= &PB{
    globalN: &n,
    grain: grain,
    bN: uint64(b.N),
    }
    body(pb)
    }()
    }
    wg.Wait()
    CQBSBMMFMJTN༬ઃҝ
    ༬ઃҝ$16ݸᏐ
    IUUQTQLHHPEFWUFTUJOH#3VO1BSBMMFM
    

    View Slide

24. View Slide

25. ུաෆඞཁతܭࢉ
    C4UPQ5JNFSC4UBSU5JNFS
    init()
    b.ResetTimer()
    for i::=0; ifoobar01()
    b.StopTimer()
    foobar02()
    b.StartTimer()
    }
    

    View Slide

26. 測試多久+測試次數
    CFODIUJNFIT CFODIUJNFY
    

    View Slide

27. 產⽣多組樣本數據
    DPVOU
    

    View Slide

28. ᒬຊᏐᎦ
    Ҡআ༗ྀٙతࢿ㘤
    

    View Slide

29. Benchmark 比較
    golang.org/x/perf/cmd/benchstat
    

    View Slide

30. 比較兩種不同版本差異
    HPUFTUWDPVOUCFODIYYYYYSVOOPOFMFYFScUFFPMEUYU
    HPUFTUWDPVOUCFODIYYYYYSVOOPOFMFYFScUFFOFXUYU
    

    View Slide

31. ᄸҰଌࢼ܈ᱪ
    benchstat old.txt
    

    View Slide

32. func (m *Metrics) computeStats() {
    /// Discard outliers.
    values ::= stats.Sample{Xs: m.Values}
    q1, q3 ::= values.Percentile(0.25), values.Percentile(0.75)
    lo, hi ::= q1-1.5*(q3-q1), q3+1.5*(q3-q1)
    for _, value ::= range m.Values {
    if lo <<= value &&& value <<= hi {
    m.RValues = append(m.RValues, value)
    }
    }
    /// Compute statistics of remaining data.
    m.Min, m.Max = stats.Bounds(m.RValues)
    m.Mean = stats.Mean(m.RValues)
    }
    ഒ࢛෼Ґڑํࣜ
    ࣗಈҠআ༗ྀٙతᏐᎦ
    golang.org/x/perf/cmd/benchstat
    

    View Slide

33. Ꮘೳલޙൺֱ
    benchstat old.txt new.txt
    

    View Slide

34. 如何找出效能瓶頸
    ࢖༻QQSPG෼ੳ$16҃هԱᱪ࢖༻ྔ
    DQVQSPpMFDQVPVUNFNQSPpMFNFNPVU
    

    View Slide

35. HPUFTUCFODIYYYYYDPVOUSVO?DQVQSPpMFDQVPVUMFYFS
    HPUPPMQQSPGDQVPVU
    

    View Slide

36. ؃ᄸҰവࣔଌࢼᏈೳ݁Ռ MJTUGVOD
    
    

    View Slide

37. ݺڣ࣍Ꮠ౷ܭ QFFLGVOD
    
    

    View Slide

38. QQSPG㗞ੜ֤छෆಉใࠂ
    

    View Slide

39. ⽤ Web 直接看比較快
    HPUPPMQQSPGIUUQDQVPVU
    

    View Slide

40. View Slide

41. 第⼀版 Parser 效能
    3FHFY1BSTFSWT-FYFS1BSTFS
    

    View Slide

42. ఏঋେ໿ഒᏈೳ
    

    View Slide

43. ఏঋେ໿ഒᏈೳ
    

    View Slide

44. 效能結果
    1BSTFS ,# .# .
    (# .)8
    
    PME
    
    
    
    
    
    VOJUTFDPOE
    ఏঋ઀ۙഒ
    DPSF(
    

    View Slide

45. /FYU5PLFOՖඅ࣌ؒ࠷ٱ
    

    View Slide

46. View Slide

47. OFX5PLFOՖඅ኷ٱ
    

    View Slide

48. func (l *Lexer) NextToken() {
    LOOP:
    switch l.char {
    case '{':
    l.token.Type = token.LeftBrace
    l.token.Literal = []byte{l.char}
    l.token.Line = l.line
    l.token.Start = l.position
    l.token.End = l.position + 1
    case ' ', '\t', '\r', '\n':
    if l.char === '\n' {
    l.line+++
    }
    l.readChar()
    goto LOOP
    

    View Slide

49. 4LJQ8IJUFTQBDF
    func (l *Lexer) NextToken() {
    l.skipWhitespace()
    switch l.char {
    case '{':
    l.token.Type = token.LeftBrace
    l.token.Literal = []byte{l.char}
    l.token.Line = l.line
    l.token.Start = l.position
    l.token.End = l.position + 1
    

    View Slide

50. 4LJQ8IJUFTQBDFमਖ਼݁Ռ
    

    View Slide

51. values ( \
    "0, 0.00014625, 0.0003375, 0.00043875, 0.000545625", \
    "0, 0.065, 0.15, 0.195, 0.2425, 0.295, 0.35, 0.42" \
    );
    values ( "0.1, 0.2, 0.3", \
    "0.11, 0.21, 0.31", \
    "0.12, 0.22, 0.32" );
    index_2 ("0.02239")
    index_2 (TVDD, 0.633)
    statetable ( " D CP SE SI " , "Q" )
    ႔ཧଟछࣈ۲ ૬Ճ
    
    

    View Slide

52. func (p *Parser) parseMultipleValue() ast.Value {
    val ::= ast.Literal{Type: ast.LiteralType}
    values ::= []string{}
    for !p.currentTokenTypeIs(token.EOF) {
    if p.currentTokenTypeIs(token.Escape) {
    p.nextToken()
    } else if p.currentTokenTypeIs(token.String) {
    values = append(values, p.parseString())
    p.nextToken()
    } else if p.currentTokenTypeIs(token.Comma) {
    values = append(values, p.parseString())
    p.nextToken()
    } else if p.currentTokenTypeIs(token.RightParentheses) {
    break
    }
    }
    val.Value = strings.Join(values, "")
    return val
    }
    

    View Slide

53. func (p *Parser) parseMultipleValue() ast.Value {
    val ::= ast.Literal{Type: ast.LiteralType}
    buf.Reset()
    for !p.currentTokenTypeIs(token.EOF) {
    if p.currentTokenTypeIs(token.Escape) {
    p.nextToken()
    } else if p.currentTokenTypeIs(token.String) {
    buf.WriteString(p.parseString())
    p.nextToken()
    } else if p.currentTokenTypeIs(token.Comma) {
    buf.WriteString(p.parseString())
    p.nextToken()
    } else if p.currentTokenTypeIs(token.RightParentheses) {
    break
    }
    }
    val.Value = buf.String()
    return val
    }
    ༻CZUFT#V⒎FSऔ୅
    

    View Slide

54. ࣈ۲႔ཧमਖ਼݁Ռ
    

    View Slide

55. values ( \
    "0, 0.00014625, 0.0003375, 0.00043875, 0.000545625", \
    "0, 0.065, 0.15, 0.195, 0.2425, 0.295, 0.35, 0.42" \
    );
    values ( "0.1, 0.2, 0.3", \
    "0.11, 0.21, 0.31", \
    "0.12, 0.22, 0.32" );
    index_2 ("0.02239")
    index_2 (TVDD, 0.633)
    statetable ( " D CP SE SI " , "Q" )
    ႔ཧଟछࣈ۲ ૬Ճ
    
    

    View Slide

56. case ast.PropertyParentheses:
    var val ast.Value
    if !p.peekTokenTypeIs(token.RightParentheses) {
    val = p.parseMultipleValue()
    prop.Value = val
    } else {
    /// single value
    /// index_2 ("0.02239")
    val = p.parseValue()
    prop.Value = val
    }
    ࣈ۲႔ཧᬓाௐ੔
    

    View Slide

57. ࣈ۲႔ཧᬓाௐ੔
    

    View Slide

58. case ',':
    l.token.Type = token.Comma
    l.token.Literal = []byte{l.char}
    l.token.Line = l.line
    l.token.Start = l.position
    l.token.End = l.position + 1
    case '"':
    l.token.Type = token.String
    l.token.Literal = l.readString()
    l.token.Line = l.line
    l.token.Start = l.position
    l.token.End = l.position + 1
    ݮগهԱᱪ࢖༻ྔ
    

    View Slide

59. case ',':
    l.token.Type = token.Comma
    l.token.Start = l.position
    l.token.End = l.position + 1
    l.token.Literal = l.Data[l.token.Start:l.token.End]
    l.token.Line = l.line
    case '"':
    l.token.Type = token.String
    l.token.Literal = l.readString()
    l.token.Line = l.line
    l.token.Start = l.position
    l.token.End = l.position + 1
    ݮগهԱᱪ࢖༻ྔ
    

    View Slide

60. ݮগهԱᱪ࢖༻ྔ
    

    View Slide

61. View Slide

62. ᩇऔძҾᥒؒతࣈ۲
    

    View Slide

63. func (l *Lexer) readString() []byte {
    l.position+++
    data ::= l.Data[l.position:]
    _, length ::= findStringLen(data)
    l.position += length
    l.readPosition = l.position + 1
    return data[:length]
    }
    ༏ԽSFBETUJOH
    

    View Slide

64. func findStringLen(data []byte) (isValid bool, length int) {
    for {
    idx ::= bytes.IndexByte(data, '"')
    if idx === -1 {
    return false, len(data)
    }
    if idx === 0 ||| (idx > 0 &&& data[idx-1] !!= '\\') {
    return true, length + idx
    }
    length += idx + 1
    data = data[idx+1:]
    }
    }
    ፙ౸ୈҰݸ݁ଋූᥒ
    

    View Slide

65. ༏Խ݁Ռ
    

    View Slide

66. ༏ԽSFBE*EFOUJpFS
    

    View Slide

67. /// check previous token
    /// example:
    /// voltage_map ( TVDD, 0.633);
    /// voltage_map (TVDD, 0.633);
    if len(l.token.Literal) > 0 &&& l.token.Literal[0] === '(' {
    l.token.Type = token.String
    l.token.Literal = l.readMapString()
    l.token.Line = l.line
    l.token.Start = l.position
    l.token.End = l.position + 1
    return
    }
    ࢿྉྔ኷େ
    ࢖༻CZUFT*OEFY#ZUF
    

    View Slide

68. ༏Խ݁Ռ
    

    View Slide

69. मਖ਼4LJQ8IJUF4QBDF'VOD
    if whitespace[l.Data[l.position]] {
    if l.Data[l.position] === '\n' {
    l.line+++
    }
    l.readChar()
    goto LOOP
    }
    var whitespace = [256]bool{
    ' ': true,
    '\t': true,
    '\n': true,
    '\r': true,
    }
    

    View Slide

70. Ꮘೳଌࢼใࠂ
    

    View Slide

71. 效能結果 (原先使⽤ 400 台機器)
    1BSTFS ,# .# .
    (# .)8
    
    PME
    
    
    
    
    
    VOJUTFDPOE
    ఏঋYYഒ
    DPSF(
    লԼ)8㑌࣍୆(3BN7.
    

    View Slide

72. Ꮘೳใࠂ -FYFS
    
    

    View Slide

73. Ꮘೳใࠂ 1BSTFS
    
    

    View Slide