Compute stddev, median and 10/90 percentile stats as well.

example/bulkstat/bulkstat.go

@@ -4,13 +4,15 @@ package main
 // in parallel.  This is useful for benchmarking purposes.
 
 import (
-	"github.com/hanwen/go-fuse/fuse"
-	"os"
+	"encoding/line"
 	"flag"
-	"time"
 	"fmt"
-	"encoding/line"
+	"github.com/hanwen/go-fuse/fuse"
+	"math"
+	"os"
 	"runtime"
+	"sort"
+	"time"
 )
 
 func main() {
@@ -41,25 +43,58 @@ func main() {
 		files = append(files, string(l))
 	}
 
-	tot := 0.0
 	totalRuns := *runs + 1
 
+	results := make([]float64, 0)
 	for j := 0; j < totalRuns; j++ {
 		result := BulkStat(*threads, files)
 		if j > 0 {
-			tot += result
+			results = append(results, result)
 		} else {
 			fmt.Println("Ignoring first run to preheat caches.")
 		}
+
 		if j < totalRuns-1 {
 			fmt.Printf("Sleeping %.2f seconds\n", *sleepTime)
 			time.Sleep(int64(*sleepTime * 1e9))
 		}
 	}
 
-	fmt.Printf("Average of %d runs: %f ms\n", *runs, tot/float64(*runs))
+	Analyze(results)
+}
+
+func Analyze(times []float64) {
+	sorted := sort.Float64Array(times)
+	sorted.Sort()
+
+	tot := 0.0
+	for _, v := range times {
+		tot += v
+	}
+	n := float64(len(times))
+
+	avg := tot / n
+	variance := 0.0
+	for _, v := range times {
+		variance += (v - avg)*(v - avg)
+	}
+	variance /= n
+
+	stddev := math.Sqrt(variance)
+
+	median := sorted[len(times)/2]
+	perc90 := sorted[int(n * 0.9)]
+	perc10 := sorted[int(n * 0.1)]
+
+	fmt.Printf(
+		"%d samples\n" +
+		"avg %.2f ms 2sigma %.2f " +
+		"median %.2fms\n"  +
+		"10%%tile %.2fms, 90%%tile %.2fms\n",
+		len(times), avg, 2*stddev, median, perc10, perc90)
 }
 
+// Returns milliseconds.
 func BulkStat(parallelism int, files []string) float64 {
 	todo := make(chan string, len(files))
 	dts := make(chan int64, parallelism)