Golang stats.Mean函數代碼示例

func calculateConfidenceInterval(s scoreResult) confInterval {
	var t0s []float64
	var t1s []float64

	// Partition the data into treatment 0 and treatment 1
	// and save the score for evaluation
	for _, each := range s.t0 {
		t0s = append(t0s, each.y)
	}

	for _, each := range s.t1 {
		t1s = append(t1s, each.y)
	}

	var ci confInterval
	//var z = 1.96 // http://www.dummies.com/how-to/content/creating-a-confidence-interval-for-the-difference-.html
	//var z = 1.645 // http://www.dummies.com/how-to/content/creating-a-confidence-interval-for-the-difference-.html
	//var z = 2.58

	var m0, _ = stats.Mean(t0s)
	var n0 = float64(len(t0s))
	var sd0, _ = stats.StandardDeviation(t0s)

	var m1, _ = stats.Mean(t1s)
	var n1 = float64(len(t1s))
	var sd1, _ = stats.StandardDeviation(t1s)

	var mDiff = m0 - m1
	var sd0s = sd0 * sd0
	var sd1s = sd1 * sd1

	ci.min = mDiff - zScore*math.Sqrt(sd1s/n1+sd0s/n0)
	ci.max = mDiff + zScore*math.Sqrt(sd1s/n1+sd0s/n0)
	ci.diff = ci.min - ci.max

	ci.t1Max = m1 + ci.max
	ci.t1Min = m1 + ci.min
	ci.diffSd = ci.diff / sd1

	// how close is the score to the middle of the confidence interval
	ci.middle = (ci.min + ci.max) / 2
	//ci.closeness = math.Abs(s.score - ci.middle)
	//ci.closeness = math.Abs(ci.diffSd - s.score)

	// Difference in sample means +- confidence interval

	//fmt.Printf("conf interval: %f to %f,  conf diff: %f, t1: %f, t1max: %f, t1min: %f, diffSd: %f\n", ci.min, ci.max, ci.diff, m1, ci.t1Max, ci.t1Min, ci.diffSd)

	return ci
}

// https://github.com/hermanschaaf/stats/blob/master/stats.go
func NormalConfidenceInterval(nums []float64) (lower float64, upper float64) {
	conf := 1.95996 // 95% confidence for the mean, http://bit.ly/Mm05eZ
	mean, _ := stats.Mean(nums)
	dev, _ := stats.StandardDeviation(nums)
	dev = dev / math.Sqrt(float64(len(nums)))
	return mean - dev*conf, mean + dev*conf
}

func bench(requests, concurrency int, images []string, args []string) {
	start := time.Now()

	timings := make([]float64, requests)
	// Create a buffered channel so our display goroutine can't slow down the workers.
	completeCh := make(chan time.Duration, requests)
	doneCh := make(chan struct{})
	current := 0
	go func() {
		for timing := range completeCh {
			timings = append(timings, timing.Seconds())
			current++
			percent := float64(current) / float64(requests) * 100
			fmt.Printf("[%3.f%%] %d/%d containers started\n", percent, current, requests)
		}
		doneCh <- struct{}{}
	}()
	session(requests, concurrency, images, args, completeCh)
	close(completeCh)
	<-doneCh

	total := time.Since(start)
	mean, _ := stats.Mean(timings)
	p90th, _ := stats.Percentile(timings, 90)
	p99th, _ := stats.Percentile(timings, 99)

	meanMillis := mean * MILLIS_IN_SECOND
	p90thMillis := p90th * MILLIS_IN_SECOND
	p99thMillis := p99th * MILLIS_IN_SECOND

	fmt.Printf("\n")
	fmt.Printf("Time taken for tests: %.3fs\n", total.Seconds())
	fmt.Printf("Time per container: %.3fms [mean] | %.3fms [90th] | %.3fms [99th]\n", meanMillis, p90thMillis, p99thMillis)
}

func updateComplexity(v *Web, pop neat.Population) {
	// Build complexity slice
	x := make([]float64, len(pop.Genomes))
	for i, g := range pop.Genomes {
		x[i] = float64(g.Complexity())
	}

	var b neat.Genome
	max := -1.0
	for _, g := range pop.Genomes {
		if g.Fitness > max {
			b = g
			max = g.Fitness
		}
	}

	// Append the record
	min, _ := stats.Min(x)
	max, _ = stats.Max(x)
	mean, _ := stats.Mean(x)

	v.complexity = append(v.complexity, [4]float64{
		min,
		mean,
		max,
		float64(b.Complexity()),
	})
}

// startStats blocks and periodically logs transaction statistics (throughput,
// success rates, durations, ...).
// TODO(tschottdorf): Use a proper metrics subsystem for this (+the store-level
// stats).
// TODO(mrtracy): Add this to TimeSeries.
func (tc *TxnCoordSender) startStats() {
	res := time.Millisecond // for duration logging resolution
	lastNow := tc.clock.PhysicalNow()
	for {
		select {
		case <-time.After(statusLogInterval):
			if !log.V(1) {
				continue
			}

			tc.Lock()
			curStats := tc.txnStats
			tc.txnStats = txnCoordStats{}
			tc.Unlock()

			now := tc.clock.PhysicalNow()

			// Tests have weird clocks.
			if now-lastNow <= 0 {
				continue
			}

			num := len(curStats.durations)
			dMax := time.Duration(stats.Max(curStats.durations))
			dMean := time.Duration(stats.Mean(curStats.durations))
			dDev := time.Duration(stats.StdDevP(curStats.durations))
			rMax := stats.Max(curStats.restarts)
			rMean := stats.Mean(curStats.restarts)
			rDev := stats.StdDevP(curStats.restarts)

			rate := float64(int64(num)*int64(time.Second)) / float64(now-lastNow)
			var pCommitted, pAbandoned, pAborted float32
			if num > 0 {
				pCommitted = 100 * float32(curStats.committed) / float32(num)
				pAbandoned = 100 * float32(curStats.abandoned) / float32(num)
				pAborted = 100 * float32(curStats.aborted) / float32(num)
			}
			log.Infof("txn coordinator: %.2f txn/sec, %.2f/%.2f/%.2f %%cmmt/abrt/abnd, %s/%s/%s avg/σ/max duration, %.1f/%.1f/%.1f avg/σ/max restarts (%d samples)",
				rate, pCommitted, pAborted, pAbandoned, util.TruncateDuration(dMean, res),
				util.TruncateDuration(dDev, res), util.TruncateDuration(dMax, res),
				rMean, rDev, rMax, num)
			lastNow = now
		case <-tc.stopper.ShouldStop():
			return
		}
	}
}

// Finalize calculation of the risk using available datapoints
func riskFinalize(op opContext, rs *slib.RRAServiceRisk) error {
	var (
		rvals []float64
		err   error
	)
	for _, x := range rs.Scenarios {
		// If the scenario had no data, don't include it in the
		// final scoring
		if x.NoData {
			continue
		}
		rvals = append(rvals, x.Score)
	}

	// Note the highest business impact value that was determined from
	// the RRA. This can be used as an indication of the business impact
	// for the service.
	rs.Risk.Impact = rs.UsedRRAAttrib.Impact
	rs.Risk.ImpactLabel, err = slib.ImpactLabelFromValue(rs.Risk.Impact)
	if err != nil {
		return err
	}

	if len(rvals) == 0 {
		// This can occur if we have no metric data, including no valid
		// information in the RRA
		logf("error in risk calculation: %q has no valid scenarios", rs.RRA.Name)
		rs.Risk.Median = 0.0
		rs.Risk.Average = 0.0
		rs.Risk.WorstCase = 0.0
		rs.Risk.MedianLabel = "unknown"
		rs.Risk.AverageLabel = "unknown"
		rs.Risk.WorstCaseLabel = "unknown"
		rs.Risk.DataClass, err = slib.DataValueFromLabel(rs.RRA.DefData)
		return nil
	}
	rs.Risk.Median, err = stats.Median(rvals)
	if err != nil {
		return err
	}
	rs.Risk.MedianLabel = slib.NormalLabelFromValue(rs.Risk.Median)
	rs.Risk.Average, err = stats.Mean(rvals)
	if err != nil {
		return err
	}
	rs.Risk.AverageLabel = slib.NormalLabelFromValue(rs.Risk.Average)
	rs.Risk.WorstCase, err = stats.Max(rvals)
	if err != nil {
		return err
	}
	rs.Risk.WorstCaseLabel = slib.NormalLabelFromValue(rs.Risk.WorstCase)

	rs.Risk.DataClass, err = slib.DataValueFromLabel(rs.RRA.DefData)
	if err != nil {
		return err
	}
	return nil
}

func (s *statistics) report() {
	for range time.Tick(time.Second) {
		s.Lock()
		writeTimes := s.writeTimes
		s.writeTimes = nil
		s.Unlock()

		// The stats functions return an error only when the input is empty.
		mean, _ := stats.Mean(writeTimes)
		stddev, _ := stats.StandardDeviation(writeTimes)
		log.Infof("wrote %d messages, latency mean=%s, stddev=%s",
			len(writeTimes), time.Duration(mean), time.Duration(stddev))
	}
}

func updateFitness(v *Web, pop neat.Population) {
	// Build fitness slice
	x := make([]float64, len(pop.Genomes))
	for i, g := range pop.Genomes {
		x[i] = g.Fitness
	}

	// Append the record
	min, _ := stats.Min(x)
	max, _ := stats.Max(x)
	mean, _ := stats.Mean(x)
	v.fitness = append(v.fitness, [3]float64{
		min,
		mean,
		max,
	})
}

// for a partition in the set of data, calculate the effective treatement
// score using (mean t1 - mean t0) / population standard deviation
func evalScore(d []coreData, rc []rowCriteria, dataSetId int) scoreResult {
	var s scoreResult
	s.rc = rc
	s.d = d
	s.score = 0
	s.dataSetId = dataSetId

	// check for minimum row threshhold
	if len(d) <= rowThreshhold {
		return s
	}

	var t0 []coreData
	var t1 []coreData
	var t0s []float64
	var t1s []float64
	var allTs []float64

	// Partition the data into treatment 0 and treatment 1
	// and save the score for evaluation
	for _, each := range d {
		// Save all responses for later SD calculation
		allTs = append(allTs, each.y)

		if each.treatment == 0 {
			t0 = append(t0, each)
			t0s = append(t0s, each.y)
		} else {
			t1 = append(t1, each)
			t1s = append(t1s, each.y)
		}
	}

	// Must have minimum threshhold of records
	if len(t0)+len(t1) < rowThreshhold {
		return s
	}

	// Must have at least one in each group
	if len(t0) == 0 || len(t1) == 0 {
		return s
	}

	// then calculate the median, also experiment with average
	var mean0, _ = stats.Mean(t0s)
	var mean1, _ = stats.Mean(t1s)
	//var meanAll, _ = stats.Mean(allTs)
	//var sd, _ = stats.StandardDeviationPopulation(allTs)

	// subtract the two t0-t1, we want t1 to be smaller
	// Note: use spooled
	// square root of ((Nt-1)St^2 + (Nc-1)Sc^2)/(Nt+Nc))
	var St, _ = stats.StandardDeviation(t1s)
	var Sc, _ = stats.StandardDeviation(t0s)
	var Nt = float64(len(t1s))
	var Nc = float64(len(t0s))

	//func calculateConfidenceInterval2(nt, nc, mt, mc, sdt, sdc float64) confInterval2
	var ci = calculateConfidenceInterval2(Nt, Nc, mean1, mean0, St, Sc)

	// If the confidence intervals overlap then not valid range
	if ci.overlap {
		return s
	}

	var St2 = St * St
	var Sc2 = Sc * Sc
	//var Ntm1 = float64(Nt - 1)
	//var Ncm1 = float64(Nc - 1)
	//var kt = Ntm1 * St2
	//var kc = Ncm1 * Sc2
	//var ksum = kt + kc
	//var Nsum = Nt + Nc
	//var sPooled = math.Sqrt(ksum / Nsum)

	//http://www.uccs.edu/~lbecker/
	var sPooled = math.Sqrt((St2 + Sc2) / 2)

	s.t0 = t0
	s.t1 = t1
	//sPooled = math.Sqrt((St2 * Sc2) / 2)

	//var _, t1confh = NormalConfidenceInterval(t1s)
	//var _, t0confh = NormalConfidenceInterval(t0s)

	//var meanValue = mean1 - mean0
	//var meanValue = (mean1/St - mean0/Sc) / sPooled

	// Score Type 1
	var meanDifference = mean1 - mean0
	//s.score = meanDifference / meanAll
	//var max, _ = stats.Max(allTs)

	//s.score = meanDifference / sPooled
	var cohensd = meanDifference / sPooled
	var a = ((Nt + Nc) * (Nt + Nc)) / (Nt + Nc)

	// Score type 5
//.........這裏部分代碼省略.........

func main() {

	d := stats.LoadRawData([]interface{}{1.1, "2", 3.0, 4, "5"})

	a, _ := stats.Min(d)
	fmt.Println(a) // 1.1

	a, _ = stats.Max(d)
	fmt.Println(a) // 5

	a, _ = stats.Sum([]float64{1.1, 2.2, 3.3})
	fmt.Println(a) // 6.6

	a, _ = stats.Mean([]float64{1, 2, 3, 4, 5})
	fmt.Println(a) // 3

	a, _ = stats.Median([]float64{1, 2, 3, 4, 5, 6, 7})
	fmt.Println(a) // 4

	m, _ := stats.Mode([]float64{5, 5, 3, 3, 4, 2, 1})
	fmt.Println(m) // [5 3]

	a, _ = stats.PopulationVariance([]float64{1, 2, 3, 4, 5})
	fmt.Println(a) // 2

	a, _ = stats.SampleVariance([]float64{1, 2, 3, 4, 5})
	fmt.Println(a) // 2.5

	a, _ = stats.MedianAbsoluteDeviationPopulation([]float64{1, 2, 3})
	fmt.Println(a) // 1

	a, _ = stats.StandardDeviationPopulation([]float64{1, 2, 3})
	fmt.Println(a) // 0.816496580927726

	a, _ = stats.StandardDeviationSample([]float64{1, 2, 3})
	fmt.Println(a) // 1

	a, _ = stats.Percentile([]float64{1, 2, 3, 4, 5}, 75)
	fmt.Println(a) // 4

	a, _ = stats.PercentileNearestRank([]float64{35, 20, 15, 40, 50}, 75)
	fmt.Println(a) // 40

	c := []stats.Coordinate{
		{1, 2.3},
		{2, 3.3},
		{3, 3.7},
		{4, 4.3},
		{5, 5.3},
	}

	r, _ := stats.LinearRegression(c)
	fmt.Println(r) // [{1 2.3800000000000026} {2 3.0800000000000014} {3 3.7800000000000002} {4 4.479999999999999} {5 5.179999999999998}]

	r, _ = stats.ExponentialRegression(c)
	fmt.Println(r) // [{1 2.5150181024736638} {2 3.032084111136781} {3 3.6554544271334493} {4 4.406984298281804} {5 5.313022222665875}]

	r, _ = stats.LogarithmicRegression(c)
	fmt.Println(r) // [{1 2.1520822363811702} {2 3.3305559222492214} {3 4.019918836568674} {4 4.509029608117273} {5 4.888413396683663}]

	s, _ := stats.Sample([]float64{0.1, 0.2, 0.3, 0.4}, 3, false)
	fmt.Println(s) // [0.2,0.4,0.3]

	s, _ = stats.Sample([]float64{0.1, 0.2, 0.3, 0.4}, 10, true)
	fmt.Println(s) // [0.2,0.2,0.4,0.1,0.2,0.4,0.3,0.2,0.2,0.1]

	q, _ := stats.Quartile([]float64{7, 15, 36, 39, 40, 41})
	fmt.Println(q) // {15 37.5 40}

	iqr, _ := stats.InterQuartileRange([]float64{102, 104, 105, 107, 108, 109, 110, 112, 115, 116, 118})
	fmt.Println(iqr) // 10

	mh, _ := stats.Midhinge([]float64{1, 3, 4, 4, 6, 6, 6, 6, 7, 7, 7, 8, 8, 9, 9, 10, 11, 12, 13})
	fmt.Println(mh) // 7.5

	tr, _ := stats.Trimean([]float64{1, 3, 4, 4, 6, 6, 6, 6, 7, 7, 7, 8, 8, 9, 9, 10, 11, 12, 13})
	fmt.Println(tr) // 7.25

	o, _ := stats.QuartileOutliers([]float64{-1000, 1, 3, 4, 4, 6, 6, 6, 6, 7, 8, 15, 18, 100})
	fmt.Printf("%+v\n", o) //  {Mild:[15 18] Extreme:[-1000 100]}

	gm, _ := stats.GeometricMean([]float64{10, 51.2, 8})
	fmt.Println(gm) // 15.999999999999991

	hm, _ := stats.HarmonicMean([]float64{1, 2, 3, 4, 5})
	fmt.Println(hm) // 2.18978102189781

	a, _ = stats.Round(2.18978102189781, 3)
	fmt.Println(a) // 2.189
}

func main() {
	verbose := flag.Bool("v", false, "verbose output")
	flag.Parse()

	file, err := os.Open("delta_data.bin")
	check(err)
	defer file.Close()

	buffer := bufio.NewReader(file)

	sizes := make([]float64, 0)
	speeds := make([]float64, 0)

	encode := qpc.NewHistory("encode")
	decode := qpc.NewHistory("decode")

	server := physics.NewState(901)
	client := physics.NewState(901)

	// initialize the base state
	for i := 0; i < 6; i += 1 {
		server.ReadNext(buffer)
		client.IncFrame()
		client.Current().Assign(server.Current())
	}

	frame := 6
	for {
		err = server.ReadNext(buffer)
		if err == io.EOF {
			break
		}
		check(err)
		frame += 1

		runtime.GC()

		// Server side
		encode.Start()
		snapshot := server.Encode()
		encode.Stop()
		// ===

		runtime.GC()

		// Client side
		decode.Start()
		client.IncFrame()
		client.Decode(snapshot)
		decode.Stop()
		// ===

		size := float64(len(snapshot)*8) / 1000.0
		sizes = append(sizes, size)

		speed := size * 60.0
		speeds = append(speeds, speed)

		equal := server.Current().Equals(client.Current())
		if *verbose {
			if !equal {
				fmt.Print("! ")
			}
			fmt.Printf("%04d %8.3fkbps %10s %10s\n", frame, speed, encode.Last(), decode.Last())
		} else {
			if equal {
				fmt.Print(".")
			} else {
				fmt.Print("X")
			}
		}
	}

	fmt.Println()
	fmt.Printf("#%d %.3fkbps ±%.3fkbps\n", len(sizes), stats.Mean(speeds), stats.StdDevS(speeds))
	fmt.Println()

	fmt.Printf("MIN %10.3f kbps\n", stats.Min(speeds))
	for _, p := range []float64{5, 10, 25, 50, 75, 90, 95} {
		fmt.Printf("P%02.f %10.3f kbps\n", p, stats.Percentile(speeds, p))
	}
	fmt.Printf("MAX %10.3f kbps\n", stats.Max(speeds))

	fmt.Println()

	fmt.Printf("TOTAL  %10.3f kb\n", stats.Sum(sizes))
	fmt.Printf("  AVG  %10.3f kb per frame\n", stats.Mean(sizes))
	fmt.Printf("  AVG  %10.3f bits per cube\n", stats.Mean(sizes)*1000/float64(len(sizes)))

	fmt.Println()
	fmt.Println("TIMING:")
	qpc.PrintSummary(encode, decode)
}

//apply transforms an array of data
func apply(data []string, transformation templates.Transformation) ([]string, []Mapping) {
	p := transformation.Parameters
	var wg sync.WaitGroup
	var mapping []Mapping

	switch transformation.Operation {
	case "toDate":
		if len(p) != 2 {
			log.Fatal("toDate transformation requires 2 parameters:  current format, new format")
		}

		oldFormat := p[0]
		newFormat := p[1]

		for i, x := range data {
			y, err := time.Parse(oldFormat, x)
			if err != nil {
				log.Print("Error parsing date with index ", i, " with format: ", oldFormat)
			} else {
				data[i] = y.Format(newFormat)
			}
		}
	case "setNull":
		for i, x := range data {
			if arrayPos(x, p) != -1 {
				data[i] = ""
			}
		}
	case "standardize":
		if len(p) != 1 {
			log.Fatal("standardize transformation requires 1 parameter:  type (min-max|z-score)")
		}

		stype := p[0]
		switch stype {
		case "min-max":
			newData := strArrToFloatArr(data)
			min, err := stats.Min(newData)
			if err != nil {
				log.Fatal("Error finding minimum of data: ", err)
			}
			max, err := stats.Max(newData)
			if err != nil {
				log.Fatal("Error finding maximum of data: ", err)
			}
			srange := max - min

			for i, x := range newData {
				data[i] = floatToString((x - min) / srange)
			}
		case "z-score":
			newData := strArrToFloatArr(data)
			mean, err := stats.Mean(newData)
			if err != nil {
				log.Fatal("Error finding mean of data: ", err)
			}
			sd, err := stats.StandardDeviation(newData)
			if err != nil {
				log.Fatal("Error finding standard deviation of data: ", err)
			}

			for i, x := range newData {
				data[i] = floatToString((x - mean) / sd)
			}
		case "decimal":
			newData := strArrToFloatArr(data)
			max, err := stats.Max(newData)
			if err != nil {
				log.Fatal("Error finding maximum of data: ", err)
			}
			min, err := stats.Min(newData)
			if err != nil {
				log.Fatal("Error finding minimum of data: ", err)
			}

			var maxAbs float64
			if math.Abs(max) > math.Abs(min) {
				maxAbs = math.Abs(max)
			} else {
				maxAbs = math.Abs(min)
			}
			c := math.Ceil(math.Log10(maxAbs))
			for i, x := range newData {
				data[i] = floatToString(x / math.Pow10(int(c)))
			}
		}
	case "binPercent":
		table := NewPivotTable(data)
		intP := strArrToIntArr(p)
		sort.Ints(intP)
		ps := NewPercentileService(*table, intP)
		mapping = ps.CreateMappings()
		ps.Bin(mapping, data)
	case "fuzzyMap":
		if len(p) != 3 {
			log.Fatal("fuzzyMap transformation requires 3 parameters:  datasource GUID, match, put")
		}

		dsGUID := p[0]
//.........這裏部分代碼省略.........

func main() {
	flag.Parse()
	n := *concurrency
	m := *total / n

	fmt.Printf("concurrency: %d\nrequests per client: %d\n\n", n, m)

	serviceMethodName := "Hello.Say"
	args := prepareArgs()

	b := make([]byte, 1024*1024)
	i, _ := args.MarshalTo(b)
	fmt.Printf("message size: %d bytes\n\n", i)

	var wg sync.WaitGroup
	wg.Add(n * m)

	var trans uint64
	var transOK uint64

	d := make([][]int64, n, n)

	//it contains warmup time but we can ignore it
	totalT := time.Now().UnixNano()
	for i := 0; i < n; i++ {
		dt := make([]int64, 0, m)
		d = append(d, dt)

		go func(i int) {
			s := &rpcx.DirectClientSelector{Network: "tcp", Address: *host}
			client := rpcx.NewClient(s)
			client.ClientCodecFunc = codec.NewProtobufClientCodec

			var reply BenchmarkMessage

			//warmup
			for j := 0; j < 5; j++ {
				client.Call(serviceMethodName, args, &reply)
			}

			for j := 0; j < m; j++ {
				t := time.Now().UnixNano()
				err := client.Call(serviceMethodName, args, &reply)
				t = time.Now().UnixNano() - t

				d[i] = append(d[i], t)

				if err == nil && reply.Field1 == "OK" {
					atomic.AddUint64(&transOK, 1)
				}

				atomic.AddUint64(&trans, 1)
				wg.Done()
			}

			client.Close()

		}(i)

	}

	wg.Wait()
	totalT = time.Now().UnixNano() - totalT
	totalT = totalT / 1000000
	fmt.Printf("took %d ms for %d requests", totalT, n*m)

	totalD := make([]int64, 0, n*m)
	for _, k := range d {
		totalD = append(totalD, k...)
	}
	totalD2 := make([]float64, 0, n*m)
	for _, k := range totalD {
		totalD2 = append(totalD2, float64(k))
	}

	mean, _ := stats.Mean(totalD2)
	median, _ := stats.Median(totalD2)
	max, _ := stats.Max(totalD2)
	min, _ := stats.Min(totalD2)

	fmt.Printf("sent     requests    : %d\n", n*m)
	fmt.Printf("received requests    : %d\n", atomic.LoadUint64(&trans))
	fmt.Printf("received requests_OK : %d\n", atomic.LoadUint64(&transOK))
	fmt.Printf("throughput  (TPS)    : %d\n", int64(n*m)*1000/totalT)
	fmt.Printf("mean: %.f ns, median: %.f ns, max: %.f ns, min: %.f ns\n", mean, median, max, min)
	fmt.Printf("mean: %d ms, median: %d ms, max: %d ms, min: %d ms\n", int64(mean/1000000), int64(median/1000000), int64(max/1000000), int64(min/1000000))

}

func summarize(vs []float64) {
	fmt.Printf("%d %.3f ±%.3f\n", len(vs), stats.Mean(vs), stats.StdDevS(vs))
}

func main() {
	flag.Parse()
	n := *concurrency
	m := *total / n

	fmt.Printf("concurrency: %d\nrequests per client: %d\n\n", n, m)

	args := prepareArgs()

	b, _ := proto.Marshal(args)
	fmt.Printf("message size: %d bytes\n\n", len(b))

	var wg sync.WaitGroup
	wg.Add(n * m)

	var trans uint64
	var transOK uint64

	d := make([][]int64, n, n)

	//it contains warmup time but we can ignore it
	totalT := time.Now().UnixNano()
	for i := 0; i < n; i++ {
		dt := make([]int64, 0, m)
		d = append(d, dt)

		go func(i int) {
			conn, err := grpc.Dial(*host, grpc.WithInsecure())
			if err != nil {
				log.Fatalf("did not connect: %v", err)
			}
			c := NewHelloClient(conn)

			//warmup
			for j := 0; j < 5; j++ {
				c.Say(context.Background(), args)
			}

			for j := 0; j < m; j++ {
				t := time.Now().UnixNano()
				reply, err := c.Say(context.Background(), args)
				t = time.Now().UnixNano() - t

				d[i] = append(d[i], t)

				if err == nil && *(reply.Field1) == "OK" {
					atomic.AddUint64(&transOK, 1)
				}

				atomic.AddUint64(&trans, 1)
				wg.Done()
			}

			conn.Close()

		}(i)

	}

	wg.Wait()
	totalT = time.Now().UnixNano() - totalT
	totalT = totalT / 1000000
	fmt.Printf("took %d ms for %d requests", totalT, n*m)

	totalD := make([]int64, 0, n*m)
	for _, k := range d {
		totalD = append(totalD, k...)
	}
	totalD2 := make([]float64, 0, n*m)
	for _, k := range totalD {
		totalD2 = append(totalD2, float64(k))
	}

	mean, _ := stats.Mean(totalD2)
	median, _ := stats.Median(totalD2)
	max, _ := stats.Max(totalD2)
	min, _ := stats.Min(totalD2)

	fmt.Printf("sent     requests    : %d\n", n*m)
	fmt.Printf("received requests    : %d\n", atomic.LoadUint64(&trans))
	fmt.Printf("received requests_OK : %d\n", atomic.LoadUint64(&transOK))
	fmt.Printf("throughput  (TPS)    : %d\n", int64(n*m)*1000/totalT)
	fmt.Printf("mean: %.f ns, median: %.f ns, max: %.f ns, min: %.f ns\n", mean, median, max, min)
	fmt.Printf("mean: %d ms, median: %d ms, max: %d ms, min: %d ms\n", int64(mean/1000000), int64(median/1000000), int64(max/1000000), int64(min/1000000))

}