seaweedfs/go/weed/benchmark.go

527 lines
14 KiB
Go
Raw Normal View History

2014-03-10 02:42:50 +00:00
package main
import (
"bufio"
"fmt"
"io"
2014-03-10 06:12:05 +00:00
"math"
"math/rand"
2014-03-10 02:42:50 +00:00
"os"
"runtime"
"runtime/pprof"
"sort"
2014-03-10 02:42:50 +00:00
"strings"
"sync"
"time"
"github.com/chrislusf/weed-fs/go/glog"
"github.com/chrislusf/weed-fs/go/operation"
"github.com/chrislusf/weed-fs/go/util"
2014-03-10 02:42:50 +00:00
)
type BenchmarkOptions struct {
2014-03-19 11:44:59 +00:00
server *string
concurrency *int
numberOfFiles *int
fileSize *int
idListFile *string
write *bool
deletePercentage *int
read *bool
sequentialRead *bool
collection *string
cpuprofile *string
maxCpu *int
2014-03-19 11:44:59 +00:00
vid2server map[string]string //cache for vid locations
2014-03-10 02:42:50 +00:00
}
var (
b BenchmarkOptions
)
func init() {
cmdBenchmark.Run = runbenchmark // break init cycle
cmdBenchmark.IsDebug = cmdBenchmark.Flag.Bool("debug", false, "verbose debug information")
b.server = cmdBenchmark.Flag.String("server", "localhost:9333", "weedfs master location")
b.concurrency = cmdBenchmark.Flag.Int("c", 16, "number of concurrent write or read processes")
2014-03-19 11:44:59 +00:00
b.fileSize = cmdBenchmark.Flag.Int("size", 1024, "simulated file size in bytes, with random(0~63) bytes padding")
2014-03-10 02:42:50 +00:00
b.numberOfFiles = cmdBenchmark.Flag.Int("n", 1024*1024, "number of files to write for each thread")
b.idListFile = cmdBenchmark.Flag.String("list", os.TempDir()+"/benchmark_list.txt", "list of uploaded file ids")
b.write = cmdBenchmark.Flag.Bool("write", true, "enable write")
2014-03-19 11:44:59 +00:00
b.deletePercentage = cmdBenchmark.Flag.Int("deletePercent", 0, "the percent of writes that are deletes")
2014-03-10 02:42:50 +00:00
b.read = cmdBenchmark.Flag.Bool("read", true, "enable read")
b.sequentialRead = cmdBenchmark.Flag.Bool("readSequentially", false, "randomly read by ids from \"-list\" specified file")
b.collection = cmdBenchmark.Flag.String("collection", "benchmark", "write data to this collection")
b.cpuprofile = cmdBenchmark.Flag.String("cpuprofile", "", "cpu profile output file")
b.maxCpu = cmdBenchmark.Flag.Int("maxCpu", 0, "maximum number of CPUs. 0 means all available CPUs")
b.vid2server = make(map[string]string)
2014-03-10 02:42:50 +00:00
}
var cmdBenchmark = &Command{
UsageLine: "benchmark -server=localhost:9333 -c=10 -n=100000",
Short: "benchmark on writing millions of files and read out",
Long: `benchmark on an empty weed file system.
2014-03-10 02:42:50 +00:00
Two tests during benchmark:
1) write lots of small files to the system
2) read the files out
2014-03-10 02:42:50 +00:00
The file content is mostly zero, but no compression is done.
2014-03-10 02:42:50 +00:00
You can choose to only benchmark read or write.
During write, the list of uploaded file ids is stored in "-list" specified file.
You can also use your own list of file ids to run read test.
2014-03-10 02:42:50 +00:00
Write speed and read speed will be collected.
The numbers are used to get a sense of the system.
2014-03-13 19:11:26 +00:00
Usually your network or the hard drive is the real bottleneck.
2014-03-13 19:11:26 +00:00
Another thing to watch is whether the volumes are evenly distributed
to each volume server. Because the 7 more benchmark volumes are randomly distributed
to servers with free slots, it's highly possible some servers have uneven amount of
benchmark volumes. To remedy this, you can use this to grow the benchmark volumes
2014-03-13 19:11:26 +00:00
before starting the benchmark command:
http://localhost:9333/vol/grow?collection=benchmark&count=5
2014-03-10 02:42:50 +00:00
After benchmarking, you can clean up the written data by deleting the benchmark collection
http://localhost:9333/col/delete?collection=benchmark
2014-03-10 02:42:50 +00:00
`,
}
var (
2014-03-19 11:44:59 +00:00
wait sync.WaitGroup
writeStats *stats
readStats *stats
serverLimitChan map[string]chan bool
2014-03-10 02:42:50 +00:00
)
2014-03-19 11:44:59 +00:00
func init() {
serverLimitChan = make(map[string]chan bool)
}
2014-03-10 02:42:50 +00:00
func runbenchmark(cmd *Command, args []string) bool {
2014-09-21 04:18:26 +00:00
fmt.Printf("This is Seaweed File System version %s %s %s\n", util.VERSION, runtime.GOOS, runtime.GOARCH)
if *b.maxCpu < 1 {
*b.maxCpu = runtime.NumCPU()
}
runtime.GOMAXPROCS(*b.maxCpu)
if *b.cpuprofile != "" {
f, err := os.Create(*b.cpuprofile)
if err != nil {
glog.Fatal(err)
}
pprof.StartCPUProfile(f)
defer pprof.StopCPUProfile()
}
2014-03-10 02:42:50 +00:00
if *b.write {
bench_write()
2014-03-10 02:42:50 +00:00
}
if *b.read {
bench_read()
2014-03-10 02:42:50 +00:00
}
return true
}
func bench_write() {
fileIdLineChan := make(chan string)
finishChan := make(chan bool)
writeStats = newStats()
idChan := make(chan int)
wait.Add(*b.concurrency)
go writeFileIds(*b.idListFile, fileIdLineChan, finishChan)
for i := 0; i < *b.concurrency; i++ {
go writeFiles(idChan, fileIdLineChan, writeStats)
}
writeStats.start = time.Now()
2014-03-19 11:44:59 +00:00
writeStats.total = *b.numberOfFiles
go writeStats.checkProgress("Writing Benchmark", finishChan)
for i := 0; i < *b.numberOfFiles; i++ {
idChan <- i
}
close(idChan)
wait.Wait()
writeStats.end = time.Now()
wait.Add(1)
finishChan <- true
finishChan <- true
close(finishChan)
wait.Wait()
writeStats.printStats()
}
func bench_read() {
fileIdLineChan := make(chan string)
finishChan := make(chan bool)
readStats = newStats()
wait.Add(*b.concurrency)
go readFileIds(*b.idListFile, fileIdLineChan)
readStats.start = time.Now()
2014-03-19 11:44:59 +00:00
readStats.total = *b.numberOfFiles
go readStats.checkProgress("Randomly Reading Benchmark", finishChan)
for i := 0; i < *b.concurrency; i++ {
go readFiles(fileIdLineChan, readStats)
}
wait.Wait()
finishChan <- true
close(finishChan)
readStats.end = time.Now()
readStats.printStats()
}
2014-03-20 20:58:56 +00:00
type delayedFile struct {
enterTime time.Time
fp *operation.FilePart
}
2014-03-10 06:12:05 +00:00
func writeFiles(idChan chan int, fileIdLineChan chan string, s *stats) {
2014-03-24 04:56:24 +00:00
delayedDeleteChan := make(chan *delayedFile, 100)
2014-03-20 20:30:34 +00:00
var waitForDeletions sync.WaitGroup
for i := 0; i < 7; i++ {
go func() {
waitForDeletions.Add(1)
2014-03-24 04:56:24 +00:00
for df := range delayedDeleteChan {
2014-03-20 20:58:56 +00:00
if df == nil {
2014-03-20 20:30:34 +00:00
break
}
2014-03-24 04:56:24 +00:00
if df.enterTime.After(time.Now()) {
time.Sleep(df.enterTime.Sub(time.Now()))
2014-03-20 20:58:56 +00:00
}
fp := df.fp
2014-03-20 20:30:34 +00:00
serverLimitChan[fp.Server] <- true
if e := util.Delete("http://" + fp.Server + "/" + fp.Fid); e == nil {
s.completed++
} else {
s.failed++
}
<-serverLimitChan[fp.Server]
}
waitForDeletions.Done()
}()
}
2014-03-10 02:42:50 +00:00
for {
if id, ok := <-idChan; ok {
start := time.Now()
2014-03-19 06:45:02 +00:00
fileSize := int64(*b.fileSize + rand.Intn(64))
fp := &operation.FilePart{Reader: &FakeReader{id: uint64(id), size: fileSize}, FileSize: fileSize}
if assignResult, err := operation.Assign(*b.server, 1, "", *b.collection, ""); err == nil {
fp.Server, fp.Fid, fp.Collection = assignResult.PublicUrl, assignResult.Fid, *b.collection
if _, ok := serverLimitChan[fp.Server]; !ok {
serverLimitChan[fp.Server] = make(chan bool, 7)
}
serverLimitChan[fp.Server] <- true
if _, err := fp.Upload(0, *b.server); err == nil {
2014-03-19 11:44:59 +00:00
if rand.Intn(100) < *b.deletePercentage {
s.total++
2014-03-24 04:56:24 +00:00
delayedDeleteChan <- &delayedFile{time.Now().Add(time.Second), fp}
2014-03-19 11:44:59 +00:00
} else {
fileIdLineChan <- fp.Fid
}
s.completed++
2014-03-19 06:45:02 +00:00
s.transferred += fileSize
} else {
s.failed++
}
2014-03-10 02:42:50 +00:00
writeStats.addSample(time.Now().Sub(start))
<-serverLimitChan[fp.Server]
2014-03-10 02:42:50 +00:00
if *cmdBenchmark.IsDebug {
fmt.Printf("writing %d file %s\n", id, fp.Fid)
}
} else {
2014-03-10 06:12:05 +00:00
s.failed++
2014-03-10 02:42:50 +00:00
println("writing file error:", err.Error())
}
} else {
break
}
}
2014-03-24 04:56:24 +00:00
close(delayedDeleteChan)
2014-03-20 20:30:34 +00:00
waitForDeletions.Wait()
2014-03-10 02:42:50 +00:00
wait.Done()
}
2014-03-10 06:12:05 +00:00
func readFiles(fileIdLineChan chan string, s *stats) {
serverLimitChan := make(map[string]chan bool)
masterLimitChan := make(chan bool, 1)
2014-03-10 02:42:50 +00:00
for {
if fid, ok := <-fileIdLineChan; ok {
if len(fid) == 0 {
continue
}
if fid[0] == '#' {
continue
}
if *cmdBenchmark.IsDebug {
fmt.Printf("reading file %s\n", fid)
}
parts := strings.SplitN(fid, ",", 2)
vid := parts[0]
start := time.Now()
if server, ok := b.vid2server[vid]; !ok {
masterLimitChan <- true
if _, now_ok := b.vid2server[vid]; !now_ok {
if ret, err := operation.Lookup(*b.server, vid); err == nil {
if len(ret.Locations) > 0 {
server = ret.Locations[0].PublicUrl
b.vid2server[vid] = server
}
2014-03-10 02:42:50 +00:00
}
}
<-masterLimitChan
2014-03-10 02:42:50 +00:00
}
if server, ok := b.vid2server[vid]; ok {
if _, ok := serverLimitChan[server]; !ok {
serverLimitChan[server] = make(chan bool, 7)
}
serverLimitChan[server] <- true
2014-03-10 02:42:50 +00:00
url := "http://" + server + "/" + fid
if bytesRead, err := util.Get(url); err == nil {
2014-03-10 06:12:05 +00:00
s.completed++
s.transferred += int64(len(bytesRead))
2014-03-10 02:42:50 +00:00
readStats.addSample(time.Now().Sub(start))
} else {
2014-03-10 06:12:05 +00:00
s.failed++
2014-03-10 02:42:50 +00:00
println("!!!! Failed to read from ", url, " !!!!!")
}
<-serverLimitChan[server]
2014-03-10 02:42:50 +00:00
} else {
2014-03-10 06:12:05 +00:00
s.failed++
2014-03-10 02:42:50 +00:00
println("!!!! volume id ", vid, " location not found!!!!!")
}
} else {
break
}
}
wait.Done()
}
func writeFileIds(fileName string, fileIdLineChan chan string, finishChan chan bool) {
file, err := os.OpenFile(fileName, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0644)
if err != nil {
glog.Fatalf("File to create file %s: %s\n", fileName, err)
}
defer file.Close()
for {
select {
case <-finishChan:
wait.Done()
return
case line := <-fileIdLineChan:
file.Write([]byte(line))
file.Write([]byte("\n"))
}
}
}
func readFileIds(fileName string, fileIdLineChan chan string) {
file, err := os.Open(fileName) // For read access.
if err != nil {
glog.Fatalf("File to read file %s: %s\n", fileName, err)
}
defer file.Close()
r := bufio.NewReader(file)
if *b.sequentialRead {
for {
if line, err := Readln(r); err == nil {
fileIdLineChan <- string(line)
} else {
break
}
}
} else {
2014-03-20 20:58:56 +00:00
lines := make([]string, 0, readStats.total)
for {
if line, err := Readln(r); err == nil {
lines = append(lines, string(line))
} else {
break
}
}
2014-03-19 11:44:59 +00:00
if len(lines) > 0 {
2014-03-20 20:58:56 +00:00
for i := 0; i < readStats.total; i++ {
2014-03-19 11:44:59 +00:00
fileIdLineChan <- lines[rand.Intn(len(lines))]
}
2014-03-10 02:42:50 +00:00
}
}
2014-03-10 02:42:50 +00:00
close(fileIdLineChan)
}
2014-03-10 06:12:05 +00:00
const (
benchResolution = 10000 //0.1 microsecond
benchBucket = 1000000000 / benchResolution
)
// An efficient statics collecting and rendering
2014-03-10 02:42:50 +00:00
type stats struct {
2014-03-10 06:12:05 +00:00
data []int
overflow []int
2014-03-10 06:12:05 +00:00
completed int
failed int
2014-03-19 11:44:59 +00:00
total int
2014-03-10 06:12:05 +00:00
transferred int64
start time.Time
end time.Time
2014-03-10 02:42:50 +00:00
}
2014-03-10 06:12:05 +00:00
var percentages = []int{50, 66, 75, 80, 90, 95, 98, 99, 100}
func newStats() *stats {
return &stats{data: make([]int, benchResolution), overflow: make([]int, 0)}
2014-03-10 02:42:50 +00:00
}
func (s *stats) addSample(d time.Duration) {
index := int(d / benchBucket)
if index < 0 {
fmt.Printf("This request takes %3.1f seconds, skipping!\n", float64(index)/10000)
} else if index < len(s.data) {
s.data[int(d/benchBucket)]++
} else {
s.overflow = append(s.overflow, index)
}
2014-03-10 02:42:50 +00:00
}
2014-03-10 06:12:05 +00:00
func (s *stats) checkProgress(testName string, finishChan chan bool) {
2014-03-10 06:12:05 +00:00
fmt.Printf("\n------------ %s ----------\n", testName)
ticker := time.Tick(time.Second)
lastCompleted, lastTransferred, lastTime := 0, int64(0), time.Now()
for {
select {
case <-finishChan:
return
case t := <-ticker:
completed, transferred, taken := s.completed-lastCompleted, s.transferred-lastTransferred, t.Sub(lastTime)
fmt.Printf("Completed %d of %d requests, %3.1f%% %3.1f/s %3.1fMB/s\n",
2014-03-19 11:44:59 +00:00
s.completed, s.total, float64(s.completed)*100/float64(s.total),
float64(completed)*float64(int64(time.Second))/float64(int64(taken)),
float64(transferred)*float64(int64(time.Second))/float64(int64(taken))/float64(1024*1024),
)
lastCompleted, lastTransferred, lastTime = s.completed, s.transferred, t
}
}
}
func (s *stats) printStats() {
2014-03-10 06:12:05 +00:00
timeTaken := float64(int64(s.end.Sub(s.start))) / 1000000000
2014-03-10 20:22:08 +00:00
fmt.Printf("\nConcurrency Level: %d\n", *b.concurrency)
2014-03-10 06:12:05 +00:00
fmt.Printf("Time taken for tests: %.3f seconds\n", timeTaken)
fmt.Printf("Complete requests: %d\n", s.completed)
fmt.Printf("Failed requests: %d\n", s.failed)
fmt.Printf("Total transferred: %d bytes\n", s.transferred)
fmt.Printf("Requests per second: %.2f [#/sec]\n", float64(s.completed)/timeTaken)
fmt.Printf("Transfer rate: %.2f [Kbytes/sec]\n", float64(s.transferred)/1024/timeTaken)
n, sum := 0, 0
min, max := 10000000, 0
for i := 0; i < len(s.data); i++ {
n += s.data[i]
sum += s.data[i] * i
if s.data[i] > 0 {
if min > i {
min = i
}
if max < i {
max = i
}
}
}
n += len(s.overflow)
for i := 0; i < len(s.overflow); i++ {
sum += s.overflow[i]
if min > s.overflow[i] {
min = s.overflow[i]
}
if max < s.overflow[i] {
max = s.overflow[i]
}
}
2014-03-10 06:12:05 +00:00
avg := float64(sum) / float64(n)
varianceSum := 0.0
for i := 0; i < len(s.data); i++ {
if s.data[i] > 0 {
d := float64(i) - avg
varianceSum += d * d * float64(s.data[i])
}
}
for i := 0; i < len(s.overflow); i++ {
d := float64(s.overflow[i]) - avg
varianceSum += d * d
}
2014-03-10 06:12:05 +00:00
std := math.Sqrt(varianceSum / float64(n))
fmt.Printf("\nConnection Times (ms)\n")
fmt.Printf(" min avg max std\n")
fmt.Printf("Total: %2.1f %3.1f %3.1f %3.1f\n", float32(min)/10, float32(avg)/10, float32(max)/10, std/10)
//printing percentiles
fmt.Printf("\nPercentage of the requests served within a certain time (ms)\n")
percentiles := make([]int, len(percentages))
for i := 0; i < len(percentages); i++ {
percentiles[i] = n * percentages[i] / 100
}
percentiles[len(percentiles)-1] = n
percentileIndex := 0
currentSum := 0
for i := 0; i < len(s.data); i++ {
currentSum += s.data[i]
if s.data[i] > 0 && percentileIndex < len(percentiles) && currentSum >= percentiles[percentileIndex] {
fmt.Printf(" %3d%% %5.1f ms\n", percentages[percentileIndex], float32(i)/10.0)
percentileIndex++
for percentileIndex < len(percentiles) && currentSum >= percentiles[percentileIndex] {
percentileIndex++
}
2014-03-10 02:42:50 +00:00
}
}
sort.Ints(s.overflow)
for i := 0; i < len(s.overflow); i++ {
currentSum++
if percentileIndex < len(percentiles) && currentSum >= percentiles[percentileIndex] {
fmt.Printf(" %3d%% %5.1f ms\n", percentages[percentileIndex], float32(s.overflow[i])/10.0)
percentileIndex++
for percentileIndex < len(percentiles) && currentSum >= percentiles[percentileIndex] {
percentileIndex++
}
}
}
2014-03-10 02:42:50 +00:00
}
// a fake reader to generate content to upload
type FakeReader struct {
id uint64 // an id number
size int64 // max bytes
}
func (l *FakeReader) Read(p []byte) (n int, err error) {
if l.size <= 0 {
return 0, io.EOF
}
if int64(len(p)) > l.size {
n = int(l.size)
} else {
n = len(p)
}
if n >= 8 {
for i := 0; i < 8; i++ {
p[i] = byte(l.id >> uint(i*8))
2014-03-10 02:42:50 +00:00
}
}
l.size -= int64(n)
return
}
func Readln(r *bufio.Reader) ([]byte, error) {
var (
isPrefix bool = true
err error = nil
line, ln []byte
)
for isPrefix && err == nil {
line, isPrefix, err = r.ReadLine()
ln = append(ln, line...)
}
return ln, err
}