Skip to content

Speed up memory db by using fixed-size keys in skiplist search #387

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 8 commits into
base: master
Choose a base branch
from
Open

Speed up memory db by using fixed-size keys in skiplist search #387

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
54c7f78
memdb: use object-oriented accessors for skiplist nodes
holiman Dec 11, 2021
667fbf8
memdb: define backing-type
holiman Dec 11, 2021
081ba6c
memdb: make tests use accessors
holiman Dec 11, 2021
de5eb3f
memdb: remove unused constants
holiman Dec 11, 2021
afed0b2
memdb: minor refactor
holiman Dec 13, 2021
825876f
stats: add separate benchmark test
holiman Dec 13, 2021
dae8fc2
memdb: add fixed-key comparison via nodeData
holiman Dec 13, 2021
80f4416
memdb: handle 32-bit platforms
holiman Dec 13, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
155 changes: 92 additions & 63 deletions leveldb/memdb/memdb.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,9 +8,12 @@
package memdb

import (
"fmt"
"math/rand"
"sync"
"unsafe"

"bytes"
"github.com/syndtr/goleveldb/leveldb/comparer"
"github.com/syndtr/goleveldb/leveldb/errors"
"github.com/syndtr/goleveldb/leveldb/iterator"
Expand All @@ -23,7 +26,10 @@ var (
ErrIterReleased = errors.New("leveldb/memdb: iterator released")
)

const tMaxHeight = 12
const (
tMaxHeight = 12 // max height of a skiplist 'tower'
branching = 4 // branching factor for the skiplist
)

type dbIter struct {
util.BasicReleaser
Expand All @@ -37,9 +43,8 @@ type dbIter struct {

func (i *dbIter) fill(checkStart, checkLimit bool) bool {
if i.node != 0 {
n := i.p.nodeData[i.node]
m := n + i.p.nodeData[i.node+nKey]
i.key = i.p.kvData[n:m]
node := i.p.nodeAt(i.node)
i.key = i.p.kvData[node.kStart():node.kEnd()]
if i.slice != nil {
switch {
case checkLimit && i.slice.Limit != nil && i.p.cmp.Compare(i.key, i.slice.Limit) >= 0:
Expand All @@ -49,7 +54,7 @@ func (i *dbIter) fill(checkStart, checkLimit bool) bool {
goto bail
}
}
i.value = i.p.kvData[m : m+i.p.nodeData[i.node+nVal]]
i.value = i.p.kvData[node.vStart():node.vEnd()]
return true
}
bail:
Expand All @@ -74,7 +79,7 @@ func (i *dbIter) First() bool {
if i.slice != nil && i.slice.Start != nil {
i.node, _ = i.p.findGE(i.slice.Start, false)
} else {
i.node = i.p.nodeData[nNext]
i.node = i.p.nodeAt(0).nextAt(0)
}
return i.fill(false, true)
}
Expand Down Expand Up @@ -127,7 +132,7 @@ func (i *dbIter) Next() bool {
i.forward = true
i.p.mu.RLock()
defer i.p.mu.RUnlock()
i.node = i.p.nodeData[i.node+nNext]
i.node = i.p.nodeAt(i.node).nextAt(0)
return i.fill(false, true)
}

Expand Down Expand Up @@ -170,36 +175,22 @@ func (i *dbIter) Release() {
}
}

const (
nKV = iota
nKey
nVal
nHeight
nNext
)

// DB is an in-memory key/value database.
type DB struct {
cmp comparer.BasicComparer
rnd *rand.Rand

mu sync.RWMutex
kvData []byte
// Node data:
// [0] : KV offset
// [1] : Key length
// [2] : Value length
// [3] : Height
// [3..height] : Next nodes
nodeData []int
cmp comparer.BasicComparer
rnd *rand.Rand
quickCmp bool

mu sync.RWMutex
kvData []byte
nodeData []nodeInt
prevNode [tMaxHeight]int
maxHeight int
n int
kvSize int
}

func (p *DB) randHeight() (h int) {
const branching = 4
h = 1
for h < tMaxHeight && p.rnd.Int()%branching == 0 {
h++
Expand All @@ -209,14 +200,31 @@ func (p *DB) randHeight() (h int) {

// Must hold RW-lock if prev == true, as it use shared prevNode slice.
func (p *DB) findGE(key []byte, prev bool) (int, bool) {
node := 0
h := p.maxHeight - 1
var (
node = 0
h = p.maxHeight - 1
qKey []byte
)
if p.quickCmp {
// If we're using quickCmp, the key we compare against needs to
// be padded to full 8 bytes.
qKey = padKey(key)
}
for {
next := p.nodeData[node+nNext+h]
next := p.nodeAt(node).nextAt(h)
cmp := 1
if next != 0 {
o := p.nodeData[next]
cmp = p.cmp.Compare(p.kvData[o:o+p.nodeData[next+nKey]], key)
o := p.nodeAt(next)
if p.quickCmp {
// If the default comparer is used, can use bytes.Compare
if cmp = o.quickCmp(qKey); cmp == 0 {
// Deep compare
cmp = bytes.Compare(p.kvData[o.kStart():o.kEnd()], key)
}
} else {
// Deep compare
cmp = p.cmp.Compare(p.kvData[o.kStart():o.kEnd()], key)
}
}
if cmp < 0 {
// Keep searching in this list
Expand All @@ -239,9 +247,9 @@ func (p *DB) findLT(key []byte) int {
node := 0
h := p.maxHeight - 1
for {
next := p.nodeData[node+nNext+h]
o := p.nodeData[next]
if next == 0 || p.cmp.Compare(p.kvData[o:o+p.nodeData[next+nKey]], key) >= 0 {
next := p.nodeAt(node).nextAt(h)
o := p.nodeAt(next)
if next == 0 || p.cmp.Compare(p.kvData[o.kStart():o.kEnd()], key) >= 0 {
if h == 0 {
break
}
Expand All @@ -257,7 +265,7 @@ func (p *DB) findLast() int {
node := 0
h := p.maxHeight - 1
for {
next := p.nodeData[node+nNext+h]
next := p.nodeAt(node).nextAt(h)
if next == 0 {
if h == 0 {
break
Expand All @@ -282,9 +290,10 @@ func (p *DB) Put(key []byte, value []byte) error {
kvOffset := len(p.kvData)
p.kvData = append(p.kvData, key...)
p.kvData = append(p.kvData, value...)
p.nodeData[node] = kvOffset
m := p.nodeData[node+nVal]
p.nodeData[node+nVal] = len(value)
// since match is exact, there's no need to set the key size again
existing := p.nodeAt(node)
m := existing.vLen()
p.nodeAt(node).setKStart(kvOffset).setVLen(len(value))
p.kvSize += len(value) - m
return nil
}
Expand All @@ -302,12 +311,13 @@ func (p *DB) Put(key []byte, value []byte) error {
p.kvData = append(p.kvData, value...)
// Node
node := len(p.nodeData)
p.nodeData = append(p.nodeData, kvOffset, len(key), len(value), h)
newN := newNode(kvOffset, len(value), h, key)
for i, n := range p.prevNode[:h] {
m := n + nNext + i
p.nodeData = append(p.nodeData, p.nodeData[m])
p.nodeData[m] = node
prev := p.nodeAt(n)
newN.setNextAt(i, prev.nextAt(i))
prev.setNextAt(i, node)
}
p.nodeData = append(p.nodeData, newN...)

p.kvSize += len(key) + len(value)
p.n++
Expand All @@ -327,13 +337,13 @@ func (p *DB) Delete(key []byte) error {
return ErrNotFound
}

h := p.nodeData[node+nHeight]
todelete := p.nodeAt(node)
h := todelete.height()
for i, n := range p.prevNode[:h] {
m := n + nNext + i
p.nodeData[m] = p.nodeData[p.nodeData[m]+nNext+i]
prev := p.nodeAt(n)
prev.setNextAt(i, todelete.nextAt(i))
}

p.kvSize -= p.nodeData[node+nKey] + p.nodeData[node+nVal]
p.kvSize -= todelete.kLen() + todelete.vLen()
p.n--
return nil
}
Expand All @@ -356,8 +366,8 @@ func (p *DB) Contains(key []byte) bool {
func (p *DB) Get(key []byte) (value []byte, err error) {
p.mu.RLock()
if node, exact := p.findGE(key, false); exact {
o := p.nodeData[node] + p.nodeData[node+nKey]
value = p.kvData[o : o+p.nodeData[node+nVal]]
n := p.nodeAt(node)
value = p.kvData[n.vStart():n.vEnd()]
} else {
err = ErrNotFound
}
Expand All @@ -374,10 +384,9 @@ func (p *DB) Get(key []byte) (value []byte, err error) {
func (p *DB) Find(key []byte) (rkey, value []byte, err error) {
p.mu.RLock()
if node, _ := p.findGE(key, false); node != 0 {
n := p.nodeData[node]
m := n + p.nodeData[node+nKey]
rkey = p.kvData[n:m]
value = p.kvData[m : m+p.nodeData[node+nVal]]
n := p.nodeAt(node)
rkey = p.kvData[n.kStart():n.kEnd()]
value = p.kvData[n.vStart():n.vEnd()]
} else {
err = ErrNotFound
}
Expand Down Expand Up @@ -446,15 +455,15 @@ func (p *DB) Reset() {
p.n = 0
p.kvSize = 0
p.kvData = p.kvData[:0]
p.nodeData = p.nodeData[:nNext+tMaxHeight]
p.nodeData[nKV] = 0
p.nodeData[nKey] = 0
p.nodeData[nVal] = 0
p.nodeData[nHeight] = tMaxHeight

p.nodeData = p.nodeData[:0]
// Add empty first element
zero := newNode(0, 0, tMaxHeight, nil)
for n := 0; n < tMaxHeight; n++ {
p.nodeData[nNext+n] = 0
zero.setNextAt(n, 0)
p.prevNode[n] = 0
}
p.nodeData = append(p.nodeData, zero...)
p.mu.Unlock()
}

Expand All @@ -469,11 +478,31 @@ func (p *DB) Reset() {
func New(cmp comparer.BasicComparer, capacity int) *DB {
p := &DB{
cmp: cmp,
quickCmp: (cmp == comparer.DefaultComparer),
Copy link
Contributor

@rjl493456442 rjl493456442 Jan 20, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's wrong here. Actually the cmp used in leveldb is iComparer with the notion of sequence and key type.
It means the quickCmp is disabled by default.

rnd: rand.New(rand.NewSource(0xdeadbeef)),
maxHeight: 1,
kvData: make([]byte, 0, capacity),
nodeData: make([]int, 4+tMaxHeight),
}
p.nodeData[nHeight] = tMaxHeight
// Add empty first element
zero := newNode(0, 0, tMaxHeight, nil)
for n := 0; n < tMaxHeight; n++ {
zero.setNextAt(n, 0)
}
p.nodeData = append(p.nodeData, zero...)
return p
}

// Stats returns some memdb runtime information.
func (p *DB) Stats() string {
p.mu.RLock()
defer p.mu.RUnlock()
dataSize := len(p.kvData)
metadataSize := len(p.nodeData) * int(unsafe.Sizeof(nodeInt(0)))
return fmt.Sprintf(`keyvalue size: %d
metadata size: %d
item count: %d
data/metadata ratio: %.02f
average kv item size: %.02f
`, dataSize, metadataSize, p.n,
float64(dataSize)/float64(metadataSize+1), float64(dataSize/p.n))
}
14 changes: 6 additions & 8 deletions leveldb/memdb/memdb_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,10 +19,9 @@ import (
func (p *DB) TestFindLT(key []byte) (rkey, value []byte, err error) {
p.mu.RLock()
if node := p.findLT(key); node != 0 {
n := p.nodeData[node]
m := n + p.nodeData[node+nKey]
rkey = p.kvData[n:m]
value = p.kvData[m : m+p.nodeData[node+nVal]]
n := p.nodeAt(node)
rkey = p.kvData[n.kStart():n.kEnd()]
value = p.kvData[n.vStart():n.vEnd()]
} else {
err = ErrNotFound
}
Expand All @@ -33,10 +32,9 @@ func (p *DB) TestFindLT(key []byte) (rkey, value []byte, err error) {
func (p *DB) TestFindLast() (rkey, value []byte, err error) {
p.mu.RLock()
if node := p.findLast(); node != 0 {
n := p.nodeData[node]
m := n + p.nodeData[node+nKey]
rkey = p.kvData[n:m]
value = p.kvData[m : m+p.nodeData[node+nVal]]
n := p.nodeAt(node)
rkey = p.kvData[n.kStart():n.kEnd()]
value = p.kvData[n.vStart():n.vEnd()]
} else {
err = ErrNotFound
}
Expand Down
Loading