Improved read/write performance by another ~2x by not calling Stat() on every read/write

README.md

@@ -97,33 +97,33 @@ Benchmarks run on a 11" Macbook with a 1.4Ghz Intel Core i7:
 ```
 $ make bench
 ...
-BenchmarkGet/128B-4         	  300000	      5144 ns/op	     400 B/op	       5 allocs/op
+BenchmarkGet/128B-4         	  300000	      5178 ns/op	     400 B/op	       5 allocs/op
-BenchmarkGet/256B-4         	  300000	      5166 ns/op	     656 B/op	       5 allocs/op
+BenchmarkGet/256B-4         	  300000	      5273 ns/op	     656 B/op	       5 allocs/op
-BenchmarkGet/512B-4         	  300000	      5284 ns/op	    1200 B/op	       5 allocs/op
+BenchmarkGet/512B-4         	  200000	      5368 ns/op	    1200 B/op	       5 allocs/op
-BenchmarkGet/1K-4           	  200000	      5779 ns/op	    2288 B/op	       5 allocs/op
+BenchmarkGet/1K-4           	  200000	      5800 ns/op	    2288 B/op	       5 allocs/op
-BenchmarkGet/2K-4           	  200000	      6396 ns/op	    4464 B/op	       5 allocs/op
+BenchmarkGet/2K-4           	  200000	      6766 ns/op	    4464 B/op	       5 allocs/op
-BenchmarkGet/4K-4           	  200000	      7716 ns/op	    9072 B/op	       5 allocs/op
+BenchmarkGet/4K-4           	  200000	      7857 ns/op	    9072 B/op	       5 allocs/op
-BenchmarkGet/8K-4           	  200000	      9802 ns/op	   17776 B/op	       5 allocs/op
+BenchmarkGet/8K-4           	  200000	      9538 ns/op	   17776 B/op	       5 allocs/op
-BenchmarkGet/16K-4          	  100000	     13299 ns/op	   34928 B/op	       5 allocs/op
+BenchmarkGet/16K-4          	  100000	     13188 ns/op	   34928 B/op	       5 allocs/op
-BenchmarkGet/32K-4          	  100000	     21819 ns/op	   73840 B/op	       5 allocs/op
+BenchmarkGet/32K-4          	  100000	     21620 ns/op	   73840 B/op	       5 allocs/op
 
-BenchmarkPut/128B-4         	  100000	     12746 ns/op	     825 B/op	       8 allocs/op
+BenchmarkPut/128B-4         	  200000	      7875 ns/op	     409 B/op	       6 allocs/op
-BenchmarkPut/256B-4         	  100000	     12937 ns/op	     954 B/op	       8 allocs/op
+BenchmarkPut/256B-4         	  200000	      8712 ns/op	     538 B/op	       6 allocs/op
-BenchmarkPut/512B-4         	  100000	     14610 ns/op	    1245 B/op	       8 allocs/op
+BenchmarkPut/512B-4         	  200000	      9832 ns/op	     829 B/op	       6 allocs/op
-BenchmarkPut/1K-4           	  100000	     16920 ns/op	    1825 B/op	       8 allocs/op
+BenchmarkPut/1K-4           	  100000	     13105 ns/op	    1410 B/op	       6 allocs/op
-BenchmarkPut/2K-4           	  100000	     22075 ns/op	    2987 B/op	       8 allocs/op
+BenchmarkPut/2K-4           	  100000	     18601 ns/op	    2572 B/op	       6 allocs/op
-BenchmarkPut/4K-4           	   30000	     40544 ns/op	    5566 B/op	       8 allocs/op
+BenchmarkPut/4K-4           	   50000	     36631 ns/op	    5151 B/op	       6 allocs/op
-BenchmarkPut/8K-4           	   20000	     63392 ns/op	   10210 B/op	       8 allocs/op
+BenchmarkPut/8K-4           	   30000	     56128 ns/op	    9798 B/op	       6 allocs/op
-BenchmarkPut/16K-4          	   10000	    108667 ns/op	   19244 B/op	       8 allocs/op
+BenchmarkPut/16K-4          	   20000	     83209 ns/op	   18834 B/op	       6 allocs/op
-BenchmarkPut/32K-4          	   10000	    129256 ns/op	   41920 B/op	       8 allocs/op
+BenchmarkPut/32K-4          	   10000	    135899 ns/op	   41517 B/op	       6 allocs/op
 
-BenchmarkScan-4             	 1000000	      1858 ns/op	     493 B/op	      25 allocs/op
+BenchmarkScan-4             	 1000000	      1851 ns/op	     493 B/op	      25 allocs/op
 ```
 
 For 128B values:
 
-* ~180,000 reads/sec
+* ~200,000 reads/sec
-*  ~80,000 writes/sec
+* ~130,000 writes/sec
 
 The full benchmark above shows linear performance as you increase key/value sizes.
 

datafile.go

@@ -1,13 +1,14 @@
 package bitcask
 
 import (
-	"errors"
 	"fmt"
 	"os"
 	"path/filepath"
 	"sync"
 	"time"
 
+	"github.com/pkg/errors"
+
 	pb "github.com/prologic/bitcask/proto"
 	"github.com/prologic/bitcask/streampb"
 )
@@ -23,11 +24,12 @@ var (
 type Datafile struct {
 	sync.RWMutex
 
-	id  int
+	id     int
-	r   *os.File
+	r      *os.File
-	w   *os.File
+	w      *os.File
-	dec *streampb.Decoder
+	offset int64
-	enc *streampb.Encoder
+	dec    *streampb.Decoder
+	enc    *streampb.Encoder
 }
 
 func NewDatafile(path string, id int, readonly bool) (*Datafile, error) {
@@ -50,16 +52,23 @@ func NewDatafile(path string, id int, readonly bool) (*Datafile, error) {
 	if err != nil {
 		return nil, err
 	}
+	stat, err := r.Stat()
+	if err != nil {
+		return nil, errors.Wrap(err, "error calling Stat()")
+	}
+
+	offset := stat.Size()
 
 	dec := streampb.NewDecoder(r)
 	enc := streampb.NewEncoder(w)
 
 	return &Datafile{
-		id:  id,
+		id:     id,
-		r:   r,
+		r:      r,
-		w:   w,
+		w:      w,
-		dec: dec,
+		offset: offset,
-		enc: enc,
+		dec:    dec,
+		enc:    enc,
 	}, nil
 }
 
@@ -87,22 +96,7 @@ func (df *Datafile) Sync() error {
 }
 
 func (df *Datafile) Size() (int64, error) {
-	var (
+	return df.offset, nil
-		stat os.FileInfo
-		err  error
-	)
-
-	if df.w == nil {
-		stat, err = df.r.Stat()
-	} else {
-		stat, err = df.w.Stat()
-	}
-
-	if err != nil {
-		return -1, err
-	}
-
-	return stat.Size(), nil
 }
 
 func (df *Datafile) Read() (e pb.Entry, err error) {
@@ -129,23 +123,17 @@ func (df *Datafile) Write(e pb.Entry) (int64, error) {
 		return -1, ErrReadonly
 	}
 
-	stat, err := df.w.Stat()
+	e.Index = df.offset
-	if err != nil {
-		return -1, err
-	}
-
-	index := stat.Size()
-
-	e.Index = index
 	e.Timestamp = time.Now().Unix()
 
 	df.Lock()
-	err = df.enc.Encode(&e)
+	n, err := df.enc.Encode(&e)
 	df.Unlock()
 
 	if err != nil {
 		return -1, err
 	}
+	df.offset += n
 
-	return index, nil
+	return e.Index, nil
 }

streampb/stream.go

@@ -28,28 +28,29 @@ type Encoder struct {
 
 // Encode takes any proto.Message and streams it to the underlying writer.
 // Messages are framed with a length prefix.
-func (e *Encoder) Encode(msg proto.Message) error {
+func (e *Encoder) Encode(msg proto.Message) (int64, error) {
 	prefixBuf := make([]byte, prefixSize)
 
 	buf, err := proto.Marshal(msg)
 	if err != nil {
-		return err
+		return 0, err
 	}
 	binary.BigEndian.PutUint64(prefixBuf, uint64(len(buf)))
 
 	if _, err := e.w.Write(prefixBuf); err != nil {
-		return errors.Wrap(err, "failed writing length prefix")
+		return 0, errors.Wrap(err, "failed writing length prefix")
 	}
 
-	if _, err = e.w.Write(buf); err != nil {
+	n, err := e.w.Write(buf)
-		return errors.Wrap(err, "failed writing marshaled data")
+	if err != nil {
+		return 0, errors.Wrap(err, "failed writing marshaled data")
 	}
 
 	if err = e.w.Flush(); err != nil {
-		return errors.Wrap(err, "failed flushing data")
+		return 0, errors.Wrap(err, "failed flushing data")
 	}
 
-	return nil
+	return int64(n + prefixSize), nil
 }
 
 // NewDecoder creates a streaming protobuf decoder.