restic/internal/repository/indexmap.go

169 lines
4.0 KiB
Go
Raw Normal View History

Chaining hash table for repository.Index These are faster to construct but slower to access. The allocation rate is halved, the peak memory usage almost halved compared to standard map. Benchmark results on linux/amd64, -benchtime=3s -count=20: name old time/op new time/op delta PackerManager-8 178ms ± 0% 178ms ± 0% ~ (p=0.231 n=20+20) DecodeIndex-8 4.54s ± 0% 4.30s ± 0% -5.20% (p=0.000 n=18+17) DecodeIndexParallel-8 4.54s ± 0% 4.30s ± 0% -5.22% (p=0.000 n=19+18) IndexHasUnknown-8 44.4ns ± 5% 50.5ns ±11% +13.82% (p=0.000 n=19+17) IndexHasKnown-8 48.3ns ± 0% 51.5ns ±12% +6.68% (p=0.001 n=16+20) IndexAlloc-8 758ms ± 1% 616ms ± 1% -18.69% (p=0.000 n=19+19) IndexAllocParallel-8 234ms ± 3% 204ms ± 2% -12.60% (p=0.000 n=20+18) MasterIndexLookupSingleIndex-8 122ns ± 0% 145ns ± 9% +18.44% (p=0.000 n=14+20) MasterIndexLookupMultipleIndex-8 369ns ± 2% 429ns ± 8% +16.27% (p=0.000 n=20+20) MasterIndexLookupSingleIndexUnknown-8 68.4ns ± 5% 74.9ns ±13% +9.47% (p=0.000 n=20+20) MasterIndexLookupMultipleIndexUnknown-8 315ns ± 3% 369ns ±11% +17.14% (p=0.000 n=20+20) MasterIndexLookupParallel/known,indices=5-8 743ns ± 1% 816ns ± 2% +9.87% (p=0.000 n=17+17) MasterIndexLookupParallel/unknown,indices=5-8 238ns ± 1% 260ns ± 2% +9.14% (p=0.000 n=19+20) MasterIndexLookupParallel/known,indices=10-8 1.01µs ± 3% 1.11µs ± 2% +9.79% (p=0.000 n=19+20) MasterIndexLookupParallel/unknown,indices=10-8 222ns ± 0% 269ns ± 2% +20.83% (p=0.000 n=16+20) MasterIndexLookupParallel/known,indices=20-8 1.06µs ± 2% 1.19µs ± 2% +12.95% (p=0.000 n=19+18) MasterIndexLookupParallel/unknown,indices=20-8 413ns ± 1% 530ns ± 1% +28.19% (p=0.000 n=18+20) SaveAndEncrypt-8 30.2ms ± 1% 30.4ms ± 0% +0.71% (p=0.000 n=19+19) LoadTree-8 540µs ± 1% 576µs ± 1% +6.73% (p=0.000 n=20+20) LoadBlob-8 5.64ms ± 0% 5.64ms ± 0% ~ (p=0.883 n=18+17) LoadAndDecrypt-8 5.93ms ± 0% 5.95ms ± 1% ~ (p=0.247 n=20+19) LoadIndex-8 25.1ms ± 0% 24.5ms ± 1% -2.54% (p=0.000 n=18+17) name old speed new speed delta PackerManager-8 296MB/s ± 0% 296MB/s ± 0% ~ (p=0.229 n=20+20) SaveAndEncrypt-8 139MB/s ± 1% 138MB/s ± 0% -0.71% (p=0.000 n=19+19) LoadBlob-8 177MB/s ± 0% 177MB/s ± 0% ~ (p=0.890 n=18+17) LoadAndDecrypt-8 169MB/s ± 0% 168MB/s ± 1% ~ (p=0.227 n=20+19) name old alloc/op new alloc/op delta PackerManager-8 91.8kB ± 0% 91.8kB ± 0% ~ (p=0.772 n=12+19) IndexAlloc-8 786MB ± 0% 400MB ± 0% -49.04% (p=0.000 n=20+18) IndexAllocParallel-8 786MB ± 0% 401MB ± 0% -49.04% (p=0.000 n=19+15) SaveAndEncrypt-8 21.0MB ± 0% 21.0MB ± 0% +0.00% (p=0.000 n=19+19) name old allocs/op new allocs/op delta PackerManager-8 1.41k ± 0% 1.41k ± 0% ~ (all equal) IndexAlloc-8 977k ± 0% 907k ± 0% -7.18% (p=0.000 n=20+20) IndexAllocParallel-8 977k ± 0% 907k ± 0% -7.17% (p=0.000 n=19+15) SaveAndEncrypt-8 73.0 ± 0% 73.0 ± 0% ~ (all equal)
2020-06-23 20:13:25 +00:00
package repository
import (
"crypto/rand"
"encoding/binary"
"github.com/restic/restic/internal/restic"
"github.com/dchest/siphash"
)
// An indexMap is a chained hash table that maps blob IDs to indexEntries.
// It allows storing multiple entries with the same key.
//
// IndexMap uses some optimizations that are not compatible with supporting
// deletions.
//
// The buckets in this hash table contain only pointers, rather than inlined
// key-value pairs like the standard Go map. This way, only a pointer array
// needs to be resized when the table grows, preventing memory usage spikes.
type indexMap struct {
// The number of buckets is always a power of two and never zero.
buckets []*indexEntry
numentries uint
key0, key1 uint64 // Key for hash randomization.
free *indexEntry // Free list.
}
const (
growthFactor = 2 // Must be a power of 2.
maxLoad = 4 // Max. number of entries per bucket.
)
// add inserts an indexEntry for the given arguments into the map,
// using id as the key.
func (m *indexMap) add(id restic.ID, packIdx int, offset, length uint32) {
switch {
case m.numentries == 0: // Lazy initialization.
m.init()
case m.numentries >= maxLoad*uint(len(m.buckets)):
m.grow()
}
h := m.hash(id)
e := m.newEntry()
e.id = id
e.next = m.buckets[h] // Prepend to existing chain.
e.packIndex = packIdx
e.offset = offset
e.length = length
m.buckets[h] = e
m.numentries++
}
// foreach calls fn for all entries in the map, until fn returns false.
func (m *indexMap) foreach(fn func(*indexEntry) bool) {
for _, e := range m.buckets {
for e != nil {
if !fn(e) {
return
}
e = e.next
}
}
}
// foreachWithID calls fn for all entries with the given id.
func (m *indexMap) foreachWithID(id restic.ID, fn func(*indexEntry)) {
if len(m.buckets) == 0 {
return
}
h := m.hash(id)
for e := m.buckets[h]; e != nil; e = e.next {
if e.id != id {
continue
}
fn(e)
}
}
// get returns the first entry for the given id.
func (m *indexMap) get(id restic.ID) *indexEntry {
if len(m.buckets) == 0 {
return nil
}
h := m.hash(id)
for e := m.buckets[h]; e != nil; e = e.next {
if e.id == id {
return e
}
}
return nil
}
func (m *indexMap) grow() {
old := m.buckets
m.buckets = make([]*indexEntry, growthFactor*len(m.buckets))
for _, e := range old {
for e != nil {
h := m.hash(e.id)
next := e.next
e.next = m.buckets[h]
m.buckets[h] = e
e = next
}
}
}
func (m *indexMap) hash(id restic.ID) uint {
// We use siphash with a randomly generated 128-bit key, to prevent
// backups of specially crafted inputs from degrading performance.
// While SHA-256 should be collision-resistant, for hash table indices
// we use only a few bits of it and finding collisions for those is
// much easier than breaking the whole algorithm.
h := uint(siphash.Hash(m.key0, m.key1, id[:]))
return h & uint(len(m.buckets)-1)
}
func (m *indexMap) init() {
const initialBuckets = 64
m.buckets = make([]*indexEntry, initialBuckets)
var buf [16]byte
if _, err := rand.Read(buf[:]); err != nil {
panic(err) // Very little we can do here.
}
m.key0 = binary.LittleEndian.Uint64(buf[:8])
m.key1 = binary.LittleEndian.Uint64(buf[8:])
}
func (m *indexMap) len() uint { return m.numentries }
func (m *indexMap) newEntry() *indexEntry {
// Allocating in batches means that we get closer to optimal space usage,
// as Go's malloc will overallocate for structures of size 56 (indexEntry
// on amd64).
//
// 256*56 and 256*48 both have minimal malloc overhead among reasonable sizes.
// See src/runtime/sizeclasses.go in the standard library.
const entryAllocBatch = 256
if m.free == nil {
free := new([entryAllocBatch]indexEntry)
for i := range free[:len(free)-1] {
free[i].next = &free[i+1]
}
m.free = &free[0]
}
e := m.free
m.free = m.free.next
return e
}
type indexEntry struct {
id restic.ID
next *indexEntry
packIndex int // Position in containing Index's packs field.
offset uint32
length uint32
}