mkbench: emit and parse write amplification in write benchmarks

To better understand the write amplification over the duration of a
write-throughput benchmark, emit the write amp after each interval.

Update `mkbench` to parse the write amplification and expose in the
output files.

Make `(*Metrics).Total()` return a `LevelMetrics` pointer, to allow
calling `WriteAmp()`, which has a pointer receiver.

cmd/pebble/write_bench.go

@@ -131,24 +131,26 @@ func initWriteBench(cmd *cobra.Command) {
 // whether the test passed or failed. Additional metadata associated with the
 // test run is also captured.
 type writeBenchResult struct {
-	name    string
-	rate    int           // The rate at which the test is currently running.
-	passed  bool          // Was the test successful at this rate.
-	elapsed time.Duration // The total elapsed time of the test.
-	bytes   uint64        // The size of the LSM.
-	levels  int           // The number of levels occupied in the LSM.
+	name     string
+	rate     int           // The rate at which the test is currently running.
+	passed   bool          // Was the test successful at this rate.
+	elapsed  time.Duration // The total elapsed time of the test.
+	bytes    uint64        // The size of the LSM.
+	levels   int           // The number of levels occupied in the LSM.
+	writeAmp float64       // The write amplification.
 }
 
 // String implements fmt.Stringer, printing a raw benchmark line. These lines
 // are used when performing analysis on a given benchmark run.
 func (r writeBenchResult) String() string {
-	return fmt.Sprintf("BenchmarkRaw%s %d ops/sec %v pass %s elapsed %d bytes %d levels",
+	return fmt.Sprintf("BenchmarkRaw%s %d ops/sec %v pass %s elapsed %d bytes %d levels %.2f writeAmp",
 		r.name,
 		r.rate,
 		r.passed,
 		r.elapsed,
 		r.bytes,
 		r.levels,
+		r.writeAmp,
 	)
 }
 
@@ -287,7 +289,7 @@ func runWriteBenchmark(_ *cobra.Command, args []string) error {
 				if writeBenchConfig.debug && i > 0 {
 					fmt.Printf("%s\n", m)
 				}
-				fmt.Println("___elapsed___clock___rate(desired)___rate(actual)___L0files___L0levels___levels______lsmBytes")
+				fmt.Println("___elapsed___clock___rate(desired)___rate(actual)___L0files___L0levels___levels______lsmBytes___writeAmp")
 			}
 
 			// Print the current stats.
@@ -300,6 +302,8 @@ func runWriteBenchmark(_ *cobra.Command, args []string) error {
 				}
 			}
 			lsmBytes := m.DiskSpaceUsage()
+			total := m.Total()
+			writeAmp := (&total).WriteAmp()
 
 			var currRate float64
 			var stalled bool
@@ -321,7 +325,7 @@ func runWriteBenchmark(_ *cobra.Command, args []string) error {
 				int(l0Sublevels) > writeBenchConfig.targetL0SubLevels
 
 			// Print the result for this tick.
-			fmt.Printf("%10s %7s %15d %14.1f %9d %10d %8d %13d\n",
+			fmt.Printf("%10s %7s %15d %14.1f %9d %10d %8d %13d %10.2f\n",
 				time.Duration(elapsed.Seconds()+0.5)*time.Second,
 				time.Duration(time.Since(clockStart).Seconds()+0.5)*time.Second,
 				desiredRate,
@@ -330,6 +334,7 @@ func runWriteBenchmark(_ *cobra.Command, args []string) error {
 				l0Sublevels,
 				nLevels,
 				lsmBytes,
+				writeAmp,
 			)
 
 			// If we're in cool-off mode, allow it to complete before resuming
@@ -351,11 +356,12 @@ func runWriteBenchmark(_ *cobra.Command, args []string) error {
 			}
 
 			r := writeBenchResult{
-				name:    name,
-				rate:    desiredRate,
-				elapsed: time.Duration(elapsed.Seconds()+0.5) * time.Second,
-				bytes:   lsmBytes,
-				levels:  nLevels,
+				name:     name,
+				rate:     desiredRate,
+				elapsed:  time.Duration(elapsed.Seconds()+0.5) * time.Second,
+				bytes:    lsmBytes,
+				levels:   nLevels,
+				writeAmp: writeAmp,
 			}
 
 			if failed {

internal/mkbench/testdata/write-throughput/summary.json

@@ -4,12 +4,14 @@
 			"name": "write/values=1024",
 			"date": "20211027",
 			"opsSec": 58825,
+			"writeAmp": 1.17,
 			"summaryPath": "20211027-pebble-write-size=1024-run_1-summary.json"
 		},
 		{
 			"name": "write/values=1024",
 			"date": "20211028",
 			"opsSec": 65525,
+			"writeAmp": 1.39,
 			"summaryPath": "20211028-pebble-write-size=1024-run_1-summary.json"
 		}
 	]

internal/mkbench/write.go

@@ -9,6 +9,7 @@ import (
 	"fmt"
 	"io"
 	"io/ioutil"
+	"math"
 	"os"
 	"path/filepath"
 	"sort"
@@ -76,7 +77,7 @@ const (
 	summaryFilename = "summary.json"
 
 	// rawRunFmt is the format string for raw benchmark data.
-	rawRunFmt = "BenchmarkRaw%s %d ops/sec %v pass %s elapsed %d bytes %d levels"
+	rawRunFmt = "BenchmarkRaw%s %d ops/sec %v pass %s elapsed %d bytes %d levels %f writeAmp"
 )
 
 func getWriteCommand() *cobra.Command {
@@ -137,13 +138,14 @@ type writePoint struct {
 	passed      bool
 	size        uint64
 	levels      int
+	writeAmp    float64
 }
 
 // formatCSV returns a comma-separated string representation of the datapoint.
 func (p writePoint) formatCSV() string {
 	return fmt.Sprintf(
-		"%d,%d,%v,%d,%d",
-		p.elapsedSecs, p.opsSec, p.passed, p.size, p.levels)
+		"%d,%d,%v,%d,%d,%.2f",
+		p.elapsedSecs, p.opsSec, p.passed, p.size, p.levels, p.writeAmp)
 }
 
 // rawWriteRun is a collection of datapoints from a single instance of a
@@ -178,6 +180,11 @@ func (r *rawWriteRun) opsPerSecSplit() int {
 	return split
 }
 
+// writeAmp returns the value of the write-amplification at the end of the run.
+func (r *rawWriteRun) writeAmp() float64 {
+	return r.points[len(r.points)-1].writeAmp
+}
+
 // formatCSV returns a comma-separated string representation of the rawWriteRun.
 // The value itself is a newline-delimited string value comprised of the CSV
 // representation of the individual writePoints.
@@ -194,10 +201,11 @@ func (r rawWriteRun) formatCSV() string {
 // value, in addition to a path to the summary.json file with the combined data
 // for the run.
 type writeRunSummary struct {
-	Name        string `json:"name"`
-	Date        string `json:"date"`
-	OpsSec      int    `json:"opsSec"`
-	SummaryPath string `json:"summaryPath"`
+	Name        string  `json:"name"`
+	Date        string  `json:"date"`
+	OpsSec      int     `json:"opsSec"`
+	WriteAmp    float64 `json:"writeAmp"`
+	SummaryPath string  `json:"summaryPath"`
 }
 
 // writeWorkloadSummary is an alias for a slice of writeRunSummaries.
@@ -232,18 +240,25 @@ func (r writeRun) summaryFilename() string {
 
 // summarize computes a writeRunSummary datapoint for the writeRun.
 func (r writeRun) summarize() writeRunSummary {
-	var sum int
+	var (
+		sumOpsSec   int
+		sumWriteAmp float64
+	)
 	for _, rr := range r.rawRuns {
-		sum += rr.opsPerSecSplit()
+		sumOpsSec += rr.opsPerSecSplit()
+		sumWriteAmp += rr.writeAmp()
 	}
+	l := len(r.rawRuns)
+
 	return writeRunSummary{
 		Name:        r.name,
 		Date:        r.date,
 		SummaryPath: r.summaryFilename(),
 		// Calculate an average across all raw runs in this run.
 		// TODO(travers): test how this works in practice, after we have
 		// gathered enough data.
-		OpsSec: sum / len(r.rawRuns),
+		OpsSec:   sumOpsSec / l,
+		WriteAmp: math.Round(100*sumWriteAmp/float64(l)) / 100, // round to 2dp.
 	}
 }
 
@@ -405,8 +420,8 @@ func (l *writeLoader) loadRaw() error {
 			var p writePoint
 			var nameInner, elapsed string
 			n, err := fmt.Sscanf(line, rawRunFmt,
-				&nameInner, &p.opsSec, &p.passed, &elapsed, &p.size, &p.levels)
-			if err != nil || n != 6 {
+				&nameInner, &p.opsSec, &p.passed, &elapsed, &p.size, &p.levels, &p.writeAmp)
+			if err != nil || n != 7 {
 				// Stumble forward on error.
 				_, _ = fmt.Fprintf(os.Stderr, "%s: %v\n", s.Text(), err)
 				continue