golang gc 優化思路以及實例分析
一個即將上線的go 寫的高頻服務，壓測的時候發現 gc 特別高，高到10%-15% 左右了，本文記錄下優化 gc 的過程和和思路。線上環境1.10.git
該文章後續仍在不斷的更新修改中， 請移步到原文地址http://dmwan.ccgithub
首先，查看gc 是否有異常，咱們可使用 gctrace 跟蹤實時的gc 。執行下面命令能夠看到gc 的實時信息。golang
GODEBUG=gctrace=1 go run cmd/agent_bin.go
輸出結果以下:redis
gc 45 @37.801s 11%: 0.19+627+0.29 ms clock, 0.38+424/621/0+0.59 ms cpu, 356->415->225 MB, 453 MB goal, 4 P gc 46 @39.126s 11%: 2.9+927+0.16 ms clock, 5.8+342/925/0+0.33 ms cpu, 361->460->275 MB, 450 MB goal, 4 P gc 47 @40.847s 12%: 0.24+1096+0.12 ms clock, 0.49+291/1007/0+0.24 ms cpu, 427->559->319 MB, 551 MB goal, 4 P gc 48 @42.771s 12%: 0.26+841+0.12 ms clock, 0.52+377/830/0+0.24 ms cpu, 486->561->271 MB, 638 MB goal, 4 P gc 49 @44.429s 12%: 3.1+890+0.40 ms clock, 6.2+492/833/0+0.81 ms cpu, 440->528->294 MB, 543 MB goal, 4 P gc 50 @46.188s 12%: 0.23+1165+0.13 ms clock, 0.47+624/1158/0+0.27 ms cpu, 471->579->323 MB, 589 MB goal, 4 P gc 51 @48.252s 13%: 0.26+1410+0.14 ms clock, 0.52+358/1336/9.9+0.28 ms cpu, 506->620->343 MB, 646 MB goal, 4 P gc 52 @50.942s 13%: 0.27+806+0.51 ms clock, 0.55+403/805/200+1.0 ms cpu, 549->657->340 MB, 687 MB goal, 4 P gc 53 @53.014s 13%: 0.10+857+0.36 ms clock, 0.21+467/851/94+0.73 ms cpu, 546->666->351 MB, 681 MB goal, 4 P
gc 45：表示第45次GC，共有4個P (線程)參與GC。json
11%: 表示gc 佔時間比。瀏覽器
0.19+627+0.29 us：STW（stop-the-world）0.19ms, 併發標記和掃描的時間627ms, STW標記的時間0.29ms。bash
0.38+424/621/0+0.59 ms cpu, 表示垃圾回收佔用cpu時間併發
356->415->225 MB, 453 MB goal,表示堆的大小，gc後堆的大小，存活堆的大小ide
453 MB goal 表示總體堆的大小爲435M。svg
根據官方描述，golang1.0 的gc 能夠降到100ms 之內，可是這裏gc 都超過1s了，這明顯是不能夠接受的，說明gc 是有很大異常的。
檢查思路，首先利用pprof 打出整個調用過程累計的堆分配圖，查出究竟是哪些模塊堆分配異常。經過代碼內嵌pprof 暴露端口的方式，終端輸出svg。
import _ "net/http/pprof"
go func() {
http.ListenAndServe("0.0.0.0:8080", nil)
}()
而後終端輸入：
go tool pprof -alloc_space -cum -svg http://127.0.0.1:8080/debug/pprof/heap > heap.svg
而後瀏覽器打開svg，找到消耗最大的調用棧：
使用-cum，是累計函數調用棧的堆分配大小，由於圖形會將調用棧很大的路線着重標識出來，這裏能夠明顯看出是本身實現的Fetch 函數的io copyBuffer 函數gc 壓力最大。這裏大體知道來源了，那究竟是咱們函數哪一個對象有問題？
接下來咱們對這個文件使用逃逸分析：直接用go build --gcflags=-m 接文件便可：
咱們的源碼是這樣的：
package httpclient
import (
"io"
"io/ioutil"
"net/http"
"net/url"
"os"
"sync"
"time"
"refresh_agent/utils"
)
var client *http.Client
var buffPool sync.Pool
func init() {
client = &http.Client{}
http.DefaultTransport.(*http.Transport).MaxIdleConnsPerHost = 2000
http.DefaultTransport.(*http.Transport).MaxIdleConns = 20000
}
type HttpClient struct{}
func NewHttpClient() *HttpClient {
httpClient := HttpClient{}
return &httpClient
}
func (this *HttpClient) replaceUrl(srcUrl string, ip string) string {
httpPrefix := "http://"
parsedUrl, err := url.Parse(srcUrl)
if err != nil {
return ""
}
return httpPrefix + ip + parsedUrl.Path
}
func (this *HttpClient) downLoadFile(resp *http.Response) error {
out, err := os.OpenFile("/dev/null", os.O_RDWR|os.O_CREATE|os.O_APPEND, 0666)
defer out.Close()
_, err = io.Copy(out, resp.Body)
return err
}
func (this *HttpClient) Fetch(dstUrl string, method string, proxyHost string, header map[string]string, preload bool, timeout int64) (*http.Response, error) {
// proxyHost 換掉 url 中請求
newUrl := this.replaceUrl(dstUrl, proxyHost)
req, _ := http.NewRequest(method, newUrl, nil)
for k, v := range header {
req.Header.Add(k, v)
}
req.Host = utils.GetUrlHost(dstUrl)
client.Timeout = time.Duration(timeout) * time.Second
resp, err := client.Do(req)
if resp == nil || err != nil {
return resp, err
}
if preload {
err := this.downLoadFile(resp)
if err != nil {
return nil, err
}
}
io.Copy(ioutil.Discard, resp.Body)
resp.Body.Close()
return resp, nil
}
逃逸分析的結果以下：
go build --gcflags=-m libs/httpclient/httpclient.go # refresh_agent/vendor/github.com/garyburd/redigo/redis gc 1 @0.078s 5%: 0.075+24+0.040 ms clock, 0.15+0.15/11/17+0.081 ms cpu, 16->16->11 MB, 42 MB goal, 2 P # refresh_agent/vendor/github.com/imroc/req gc 1 @0.071s 3%: 0.017+26+0.050 ms clock, 0.034+0.22/6.0/22+0.10 ms cpu, 16->17->11 MB, 42 MB goal, 2 P # refresh_agent/vendor/github.com/BurntSushi/toml gc 1 @0.054s 6%: 0.086+43+0.062 ms clock, 0.17+0.14/12/14+0.12 ms cpu, 16->17->12 MB, 42 MB goal, 2 P # refresh_agent/vendor/github.com/modern-go/reflect2 gc 1 @0.054s 7%: 0.070+28+0.051 ms clock, 0.14+0.15/12/14+0.10 ms cpu, 16->17->10 MB, 42 MB goal, 2 P gc 2 @0.334s 3%: 0.016+33+0.055 ms clock, 0.032+0.54/12/20+0.11 ms cpu, 37->38->14 MB, 54 MB goal, 2 P # refresh_agent/vendor/github.com/json-iterator/go gc 1 @0.045s 5%: 0.068+27+0.040 ms clock, 0.13+2.5/4.7/16+0.081 ms cpu, 16->16->11 MB, 42 MB goal, 2 P gc 2 @0.211s 9%: 0.029+107+0.12 ms clock, 0.058+0.96/52/0+0.24 ms cpu, 39->44->26 MB, 58 MB goal, 2 P # command-line-arguments libs/httpclient/httpclient.go:18:6: can inline init.0 libs/httpclient/httpclient.go:26:6: can inline NewHttpClient libs/httpclient/httpclient.go:19:24: &http.Client literal escapes to heap libs/httpclient/httpclient.go:28:9: &httpClient escapes to heap libs/httpclient/httpclient.go:27:2: moved to heap: httpClient libs/httpclient/httpclient.go:31:62: leaking param: srcUrl libs/httpclient/httpclient.go:37:25: httpPrefix + ip + parsedUrl.Path escapes to heap libs/httpclient/httpclient.go:31:62: (*HttpClient).replaceUrl this does not escape libs/httpclient/httpclient.go:31:62: (*HttpClient).replaceUrl ip does not escape libs/httpclient/httpclient.go:44:18: out escapes to heap libs/httpclient/httpclient.go:44:28: resp.Body escapes to heap libs/httpclient/httpclient.go:41:59: leaking param content: resp libs/httpclient/httpclient.go:41:59: (*HttpClient).downLoadFile this does not escape libs/httpclient/httpclient.go:48:151: leaking param: dstUrl libs/httpclient/httpclient.go:48:151: leaking param: method libs/httpclient/httpclient.go:48:151: leaking param content: header libs/httpclient/httpclient.go:69:30: resp.Body escapes to heap libs/httpclient/httpclient.go:48:151: (*HttpClient).Fetch this does not escape libs/httpclient/httpclient.go:48:151: (*HttpClient).Fetch proxyHost does not escape <autogenerated>:1:0: leaking param: .anon0 <autogenerated>:1:0: leaking param: .this <autogenerated>:1:0: leaking param: io.p <autogenerated>:1:0: leaking param: .this <autogenerated>:1:0: os.(*File).close .this does not escape <autogenerated>:1:0: leaking param: .this <autogenerated>:1:0: leaking param: .this <autogenerated>:1:0: leaking param: .this <autogenerated>:1:0: leaking param: context.key <autogenerated>:1:0: leaking param: .this
這裏咱們發現resp.Body 逃逸到了堆上。和咱們原始的但願是不一致的，咱們的目的，是但願將resp.body 重定向到空設備中去，這裏的對象最好是分配到棧區。這裏看不出，到底哪一步反覆建立對象了。
繼續追io.CopyBuffer,查看io.Copy 的源碼：
func Copy(dst Writer, src Reader) (written int64, err error) {
return copyBuffer(dst, src, nil)
}
// CopyBuffer is identical to Copy except that it stages through the
// provided buffer (if one is required) rather than allocating a
// temporary one. If buf is nil, one is allocated; otherwise if it has
// zero length, CopyBuffer panics.
func CopyBuffer(dst Writer, src Reader, buf []byte) (written int64, err error) {
if buf != nil && len(buf) == 0 {
panic("empty buffer in io.CopyBuffer")
}
return copyBuffer(dst, src, buf)
}
func copyBuffer(dst Writer, src Reader, buf []byte) (written int64, err error) {
// If the reader has a WriteTo method, use it to do the copy.
// Avoids an allocation and a copy.
if wt, ok := src.(WriterTo); ok {
return wt.WriteTo(dst)
}
// Similarly, if the writer has a ReadFrom method, use it to do the copy.
if rt, ok := dst.(ReaderFrom); ok {
return rt.ReadFrom(src)
}
size := 32 * 1024
if l, ok := src.(*LimitedReader); ok && int64(size) > l.N {
if l.N < 1 {
size = 1
} else {
size = int(l.N)
}
}
if buf == nil {
buf = make([]byte, size) #這裏反覆建立 大小爲32k的slice
}
看到調用棧提示的函數，發現，咱們調用copy 的時候，若是不傳入buffer， 內部會反覆建立這個buf。基本就定位到問題了。
那麼咱們知道某個對象會一直被gc 回收，怎麼處理呢，通用作法是，使用對象池。
通過改造，咱們copy 的時候傳一個buffer 進去，這個buffer 從sync.Pool 中去取：
var buffPool sync.Pool
func (this *HttpClient) downLoadFile(resp *http.Response) error {
//err write /dev/null: bad file descriptor#
out, err := os.OpenFile("/dev/null", os.O_RDWR|os.O_CREATE|os.O_APPEND, 0666)
defer out.Close()
buf := buffPool.Get()
if buf == nil {
buf = make([]byte, 32*1024)
}
io.CopyBuffer(out, resp.Body, buf.([]byte))
buffPool.Put(buf)
return err
使用這種方式後，這裏的堆佔用大小立馬降下來了：
經過這樣的思路，將堆累計分配比較頻繁的一些點所有找出來後，gc降到了100ms 如下。
記錄幾個gc 優化的點, 高頻請求必定要注意：
1，函數儘可能不要返回map， slice對象, 這種頻繁調用的函數會給gc 帶來壓力。
2，小對象要合併。
3，函數頻繁建立的簡單的對象，直接返回對象，效果比返回指針效果要好。
4，避不開，能用sync.Pool 就用，雖然有人說1.10 後不推薦使用sync.Pool，可是壓測來看，確實仍是用效果，堆累計分配大小能減小一半以上。
5，類型轉換要注意，官方用法消耗特別大，推薦使用雨痕的方式。
6，避免反覆建立slice。
相關文章
- 1. 最佳路徑分析實現思路及優化研究
- 2. Unity GC以及優化（轉）
- 3. hbase實時優化思路
- 4. golang gc
- 5. 【4】JVM-GC設計思路分析
- 6. android 內存優化-MAT的使用及實例分析
- 7. Unity中的GC以及優化
- 8. 淺談Unity中的GC以及優化
- 9. Spark性能優化系列：內存調優思路分析
- 10. golang rpc實例分析
- 更多相關文章...
- • SEO - 搜索引擎優化 - 網站建設指南
- • Spring實例化Bean的三種方法 - Spring教程
- • Flink 數據傳輸及反壓詳解
- • 算法總結-廣度優先算法
相關標籤/搜索
每日一句
-
每一个你不满意的现在，都有一个你没有努力的曾经。
歡迎關注本站公眾號,獲取更多信息
相關文章
- 1. 最佳路徑分析實現思路及優化研究
- 2. Unity GC以及優化（轉）
- 3. hbase實時優化思路
- 4. golang gc
- 5. 【4】JVM-GC設計思路分析
- 6. android 內存優化-MAT的使用及實例分析
- 7. Unity中的GC以及優化
- 8. 淺談Unity中的GC以及優化
- 9. Spark性能優化系列：內存調優思路分析
- 10. golang rpc實例分析