Improve Get/Put performance with optional mempooling (#36)

* avoid unnecessary use of encoder/decoder to decrease memory allocations

* add an optional configurable mempool to avoid extra allocs

* add doc.go with examples

README.md

@@ -4,14 +4,9 @@
 [![CodeCov](https://codecov.io/gh/prologic/bitcask/branch/master/graph/badge.svg)](https://codecov.io/gh/prologic/bitcask)
 [![Go Report Card](https://goreportcard.com/badge/prologic/bitcask)](https://goreportcard.com/report/prologic/bitcask)
 [![GoDoc](https://godoc.org/github.com/prologic/bitcask?status.svg)](https://godoc.org/github.com/prologic/bitcask) 
-[![Sourcegraph](https://sourcegraph.com/github.com/prologic/bitcask/-/badge.svg)](https://sourcegraph.com/github.com/prologic/bitcask?badge)
-[![Github all releases](https://img.shields.io/github/downloads/prologic/bitcask/total.svg)](https://github.com/prologic/bitcask/releases)
 [![GitHub license](https://img.shields.io/github/license/prologic/bitcask.svg)](https://github.com/prologic/bitcask)
 
-[![](https://images.microbadger.com/badges/version/prologic/bitcask.svg)](https://microbadger.com/images/prologic/bitcask)
-[![](https://images.microbadger.com/badges/image/prologic/bitcask.svg)](https://microbadger.com/images/prologic/bitcask)
-
-A high performance Key/Value store written in [Go](https://golang.org) with a predictable read/write performance and high throughput. Uses a [Bitcask](https://en.wikipedia.org/wiki/Bitcask) on-disk layout (LSM+WAL) similar to [Riak](https://riak.com/). 🗃️
+A high performance Key/Value store written in [Go](https://golang.org) with a predictable read/write performance and high throughput. Uses a [Bitcask](https://en.wikipedia.org/wiki/Bitcask) on-disk layout (LSM+WAL) similar to [Riak](https://riak.com/)
 
 For a more feature-complete Redis-compatible server, distributed key/value store have a look at [Bitraft](https://github.com/prologic/bitraft) which uses this library as its backend. Use [Bitcask](https://github.com/prologic/bitcask) as a starting point or if you want to embed in your application, use [Bitraft](https://github.com/prologic/bitraft) if you need a complete server/client solution with high availability with a Redis-compatible API.
 
@@ -21,14 +16,14 @@ For a more feature-complete Redis-compatible server, distributed key/value store
 * Builtin CLI (`bitcask`)
 * Builtin Redis-compatible server (`bitcaskd`)
 * Predictable read/write performance
-* Low latecny
+* Low latency
 * High throughput (See: [Performance](README.md#Performance) )
 
 ## Development
 
 1. Get the source
 
-```#!bash
+```#!sh
 $ git clone https://github.com/prologic/bitcask.git
 ```
 
@@ -37,20 +32,19 @@ $ git clone https://github.com/prologic/bitcask.git
 This library uses [Protobuf](https://github.com/protocolbuffers/protobuf) to serialize data on disk. Please follow the
 instructions for installing `protobuf` on your system. You will also need the
 following Go libraries/tools to generate Go code from Protobuf defs:
+- [protoc-gen-go](https://github.com/golang/protobuf)
 
 3. Build the project
 
-```#!bash
+```#!sh
 $ make
 ```
 
 This will invoke `go generate` and `go build`.
 
-- [protoc-gen-go](https://github.com/golang/protobuf)
-
 ## Install
 
-```#!bash
+```#!sh
 $ go get github.com/prologic/bitcask
 ```
 
@@ -58,7 +52,7 @@ $ go get github.com/prologic/bitcask
 
 Install the package into your project:
 
-```#!bash
+```#!sh
 $ go get github.com/prologic/bitcask
 ```
 
@@ -80,7 +74,7 @@ documentation and other examples.
 
 ## Usage (tool)
 
-```#!bash
+```#!sh
 $ bitcask -p /tmp/db set Hello World
 $ bitcask -p /tmp/db get Hello
 World
@@ -90,14 +84,14 @@ World
 
 There is also a builtin very  simple Redis-compatible server called `bitcaskd`:
 
-```#!bash
+```#!sh
 $ ./bitcaskd ./tmp
 INFO[0000] starting bitcaskd v0.0.7@146f777              bind=":6379" path=./tmp
 ```
 
 Example session:
 
-```
+```#!sh
 $ telnet localhost 6379
 Trying ::1...
 Connected to localhost.
@@ -122,7 +116,7 @@ Connection closed by foreign host.
 
 You can also use the [Bitcask Docker Image](https://cloud.docker.com/u/prologic/repository/docker/prologic/bitcask):
 
-```#!bash
+```#!sh
 $ docker pull prologic/bitcask
 $ docker run -d -p 6379:6379 prologic/bitcask
 ```
@@ -131,7 +125,7 @@ $ docker run -d -p 6379:6379 prologic/bitcask
 
 Benchmarks run on a 11" Macbook with a 1.4Ghz Intel Core i7:
 
-```
+```#!sh
 $ make bench
 ...
 BenchmarkGet/128B-4         	  300000	      4071 ns/op	  31.43 MB/s	     608 B/op	       7 allocs/op

bitcask.go

@@ -35,6 +35,10 @@ var (
 	// ErrDatabaseLocked is the error returned if the database is locked
 	// (typically opened by another process)
 	ErrDatabaseLocked = errors.New("error: database locked")
+
+	// ErrCreatingMemPool is the error returned when trying to configurate
+	// the mempool fails
+	ErrCreatingMemPool = errors.New("error: creating the mempool failed")
 )
 
 // Bitcask is a struct that represents a on-disk LSM and WAL data structure
@@ -420,6 +424,8 @@ func Open(path string, options ...Option) (*Bitcask, error) {
 		}
 	}
 
+	internal.ConfigureMemPool(bitcask.config.maxConcurrency)
+
 	locked, err := bitcask.Flock.TryLock()
 	if err != nil {
 		return nil, err

bitcask_test.go

@@ -468,8 +468,9 @@ func TestLocking(t *testing.T) {
 }
 
 type benchmarkTestCase struct {
-	name string
-	size int
+	name     string
+	size     int
+	withPool bool
 }
 
 func BenchmarkGet(b *testing.B) {
@@ -484,22 +485,25 @@ func BenchmarkGet(b *testing.B) {
 	}
 	defer os.RemoveAll(testdir)
 
-	db, err := Open(testdir)
-	if err != nil {
-		b.Fatal(err)
-	}
-	defer db.Close()
-
 	tests := []benchmarkTestCase{
-		{"128B", 128},
-		{"256B", 256},
-		{"512B", 512},
-		{"1K", 1024},
-		{"2K", 2048},
-		{"4K", 4096},
-		{"8K", 8192},
-		{"16K", 16384},
-		{"32K", 32768},
+		{"128B", 128, false},
+		{"128BWithPool", 128, true},
+		{"256B", 256, false},
+		{"256BWithPool", 256, true},
+		{"512B", 512, false},
+		{"512BWithPool", 512, true},
+		{"1K", 1024, false},
+		{"1KWithPool", 1024, true},
+		{"2K", 2048, false},
+		{"2KWithPool", 2048, true},
+		{"4K", 4096, false},
+		{"4KWithPool", 4096, true},
+		{"8K", 8192, false},
+		{"8KWithPool", 8192, true},
+		{"16K", 16384, false},
+		{"16KWithPool", 16384, true},
+		{"32K", 32768, false},
+		{"32KWithPool", 32768, true},
 	}
 
 	for _, tt := range tests {
@@ -509,6 +513,18 @@ func BenchmarkGet(b *testing.B) {
 			key := "foo"
 			value := []byte(strings.Repeat(" ", tt.size))
 
+			options := []Option{
+				WithMaxKeySize(len(key)),
+				WithMaxValueSize(tt.size),
+			}
+			if tt.withPool {
+				options = append(options, WithMemPool(1))
+			}
+			db, err := Open(testdir, options...)
+			if err != nil {
+				b.Fatal(err)
+			}
+
 			err = db.Put(key, value)
 			if err != nil {
 				b.Fatal(err)
@@ -524,6 +540,8 @@ func BenchmarkGet(b *testing.B) {
 					b.Errorf("unexpected value")
 				}
 			}
+			b.StopTimer()
+			db.Close()
 		})
 	}
 }
@@ -547,15 +565,15 @@ func BenchmarkPut(b *testing.B) {
 	defer db.Close()
 
 	tests := []benchmarkTestCase{
-		{"128B", 128},
-		{"256B", 256},
-		{"512B", 512},
-		{"1K", 1024},
-		{"2K", 2048},
-		{"4K", 4096},
-		{"8K", 8192},
-		{"16K", 16384},
-		{"32K", 32768},
+		{"128B", 128, false},
+		{"256B", 256, false},
+		{"512B", 512, false},
+		{"1K", 1024, false},
+		{"2K", 2048, false},
+		{"4K", 4096, false},
+		{"8K", 8192, false},
+		{"16K", 16384, false},
+		{"32K", 32768, false},
 	}
 
 	for _, tt := range tests {

doc.go

@@ -0,0 +1,13 @@
+// Package bitcask implements a high-performance key-value store based on a
+// WAL and LSM.
+//
+// By default, the client assumes a default configuration regarding maximum key size,
+// maximum value size, maximum datafile size, and memory pools to avoid allocations.
+// Refer to Constants section to know default values.
+//
+// For extra performance, configure the memory pool option properly. This option
+// requires to specify the maximum number of concurrent use of the package. Failing to
+// set a high-enough value would impact latency and throughput. Likewise, overestimating
+// would yield in an unnecessary big memory footprint.
+// The default configuration doesn't use a memory pool.
+package bitcask

doc_test.go

@@ -0,0 +1,14 @@
+package bitcask
+
+func Example() {
+	_, _ = Open("path/to/db")
+}
+
+func Example_withOptions() {
+	opts := []Option{
+		WithMaxKeySize(1024),
+		WithMaxValueSize(4096),
+		WithMemPool(10),
+	}
+	_, _ = Open("path/to/db", opts...)
+}

go.mod

@@ -6,6 +6,7 @@ require (
 	github.com/golang/protobuf v1.3.2
 	github.com/konsorten/go-windows-terminal-sequences v1.0.2 // indirect
 	github.com/magiconair/properties v1.8.1 // indirect
+	github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c
 	github.com/pelletier/go-toml v1.4.0 // indirect
 	github.com/pkg/errors v0.8.1
 	github.com/prologic/trie v0.0.0-20190322091023-3972df81f9b5

go.sum

@@ -77,6 +77,8 @@ github.com/mitchellh/mapstructure v1.1.2 h1:fmNYVwqnSfB9mZU6OS2O6GsXM+wcskZDuKQz
 github.com/mitchellh/mapstructure v1.1.2/go.mod h1:FVVH3fgwuzCH5S8UJGiWEs2h04kUh9fWfEaFds41c1Y=
 github.com/mwitkow/go-conntrack v0.0.0-20161129095857-cc309e4a2223/go.mod h1:qRWi+5nqEBWmkhHvq77mSJWrCKwh8bxhgT7d/eI7P4U=
 github.com/oklog/ulid v1.3.1/go.mod h1:CirwcVhetQ6Lv90oh/F+FBtV6XMibvdAFo93nm5qn4U=
+github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c h1:rp5dCmg/yLR3mgFuSOe4oEnDDmGLROTvMragMUXpTQw=
+github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c/go.mod h1:X07ZCGwUbLaax7L0S3Tw4hpejzu63ZrrQiUe6W0hcy0=
 github.com/pelletier/go-toml v1.2.0 h1:T5zMGML61Wp+FlcbWjRDT7yAxhJNAiPPLOFECq181zc=
 github.com/pelletier/go-toml v1.2.0/go.mod h1:5z9KED0ma1S8pY6P1sdut58dfprrGBbd/94hg7ilaic=
 github.com/pelletier/go-toml v1.4.0 h1:u3Z1r+oOXJIkxqw34zVhyPgjBsm6X2wn21NWs/HfSeg=

internal/datafile.go

@@ -1,26 +1,31 @@
 package internal
 
 import (
-	"bytes"
 	"fmt"
 	"os"
 	"path/filepath"
 	"sync"
 
+	"github.com/oxtoacart/bpool"
 	"github.com/pkg/errors"
 	"golang.org/x/exp/mmap"
 
+	"github.com/gogo/protobuf/proto"
 	pb "github.com/prologic/bitcask/internal/proto"
 	"github.com/prologic/bitcask/internal/streampb"
 )
 
 const (
 	DefaultDatafileFilename = "%09d.data"
+	prefixSize              = 8
 )
 
 var (
 	ErrReadonly  = errors.New("error: read only datafile")
 	ErrReadError = errors.New("error: read error")
+
+	memPool   *bpool.BufferPool
+	mxMemPool sync.RWMutex
 )
 
 type Datafile struct {
@@ -136,7 +141,17 @@ func (df *Datafile) Read() (e pb.Entry, n int64, err error) {
 func (df *Datafile) ReadAt(index, size int64) (e pb.Entry, err error) {
 	var n int
 
-	b := make([]byte, size)
+	var b []byte
+	if memPool == nil {
+		b = make([]byte, size)
+	} else {
+		poolSlice := memPool.Get()
+		if poolSlice.Cap() < int(size) {
+			poolSlice.Grow(int(size) - poolSlice.Cap())
+		}
+		defer memPool.Put(poolSlice)
+		b = poolSlice.Bytes()[:size]
+	}
 
 	if df.w == nil {
 		n, err = df.ra.ReadAt(b, index)
@@ -151,9 +166,10 @@ func (df *Datafile) ReadAt(index, size int64) (e pb.Entry, err error) {
 		return
 	}
 
-	buf := bytes.NewBuffer(b)
-	dec := streampb.NewDecoder(buf)
-	_, err = dec.Decode(&e)
+	err = proto.Unmarshal(b[prefixSize:], &e)
+	if err != nil {
+		return
+	}
 	return
 }
 
@@ -175,3 +191,15 @@ func (df *Datafile) Write(e pb.Entry) (int64, int64, error) {
 
 	return e.Offset, n, nil
 }
+
+// ConfigureMemPool configurate the mempool accordingly
+func ConfigureMemPool(maxConcurrency *int) {
+	mxMemPool.Lock()
+	defer mxMemPool.Unlock()
+	if maxConcurrency == nil {
+		memPool = nil
+	} else {
+		memPool = bpool.NewBufferPool(*maxConcurrency)
+	}
+	return
+}

internal/keydir.go

@@ -42,17 +42,15 @@ func (k *Keydir) Add(key string, fileid int, offset, size int64) Item {
 
 func (k *Keydir) Get(key string) (Item, bool) {
 	k.RLock()
-	defer k.RUnlock()
-
 	item, ok := k.kv[key]
+	k.RUnlock()
 	return item, ok
 }
 
 func (k *Keydir) Delete(key string) {
 	k.Lock()
-	defer k.Unlock()
-
 	delete(k.kv, key)
+	k.Unlock()
 }
 
 func (k *Keydir) Len() int {
@@ -63,11 +61,11 @@ func (k *Keydir) Keys() chan string {
 	ch := make(chan string)
 	go func() {
 		k.RLock()
-		defer k.RUnlock()
 		for key := range k.kv {
 			ch <- key
 		}
 		close(ch)
+		k.RUnlock()
 	}()
 	return ch
 }

options.go

@@ -1,5 +1,7 @@
 package bitcask
 
+import "errors"
+
 const (
 	// DefaultMaxDatafileSize is the default maximum datafile size in bytes
 	DefaultMaxDatafileSize = 1 << 20 // 1MB
@@ -11,6 +13,12 @@ const (
 	DefaultMaxValueSize = 1 << 16 // 65KB
 )
 
+var (
+	// ErrMaxConcurrencyLowerEqZero is the error returned for
+	// maxConcurrency option not greater than zero
+	ErrMaxConcurrencyLowerEqZero = errors.New("error: maxConcurrency must be greater than zero")
+)
+
 // Option is a function that takes a config struct and modifies it
 type Option func(*config) error
 
@@ -18,6 +26,7 @@ type config struct {
 	maxDatafileSize int
 	maxKeySize      int
 	maxValueSize    int
+	maxConcurrency  *int
 }
 
 func newDefaultConfig() *config {
@@ -51,3 +60,14 @@ func WithMaxValueSize(size int) Option {
 		return nil
 	}
 }
+
+// WithMemPool configures usage of a memory pool to avoid allocations
+func WithMemPool(maxConcurrency int) Option {
+	return func(cfg *config) error {
+		if maxConcurrency <= 0 {
+			return ErrMaxConcurrencyLowerEqZero
+		}
+		cfg.maxConcurrency = &maxConcurrency
+		return nil
+	}
+}