Replace sharded hash function/misc mods
This commit is contained in:
parent
8a2f4f19ef
commit
03284ca422
|
@ -1472,7 +1472,8 @@ func BenchmarkRWMutexMapGetConcurrent(b *testing.B) {
|
|||
|
||||
func BenchmarkCacheGetManyConcurrent(b *testing.B) {
|
||||
// This is the same as BenchmarkCacheGetConcurrent, but its result
|
||||
// can be compared against BenchmarkShardedCacheGetManyConcurrent.
|
||||
// can be compared against BenchmarkShardedCacheGetManyConcurrent
|
||||
// in sharded_test.go.
|
||||
b.StopTimer()
|
||||
n := 10000
|
||||
tc := New(DefaultExpiration, 0)
|
||||
|
@ -1497,31 +1498,6 @@ func BenchmarkCacheGetManyConcurrent(b *testing.B) {
|
|||
wg.Wait()
|
||||
}
|
||||
|
||||
func BenchmarkShardedCacheGetManyConcurrent(b *testing.B) {
|
||||
b.StopTimer()
|
||||
n := 10000
|
||||
tsc := unexportedNewSharded(20, DefaultExpiration, 0)
|
||||
keys := make([]string, n)
|
||||
for i := 0; i < n; i++ {
|
||||
k := "foo" + strconv.Itoa(n)
|
||||
keys[i] = k
|
||||
tsc.Set(k, "bar", DefaultExpiration)
|
||||
}
|
||||
each := b.N / n
|
||||
wg := new(sync.WaitGroup)
|
||||
wg.Add(n)
|
||||
for _, v := range keys {
|
||||
go func() {
|
||||
for j := 0; j < each; j++ {
|
||||
tsc.Get(v)
|
||||
}
|
||||
wg.Done()
|
||||
}()
|
||||
}
|
||||
b.StartTimer()
|
||||
wg.Wait()
|
||||
}
|
||||
|
||||
func BenchmarkCacheSet(b *testing.B) {
|
||||
b.StopTimer()
|
||||
tc := New(DefaultExpiration, 0)
|
||||
|
|
77
sharded.go
77
sharded.go
|
@ -1,8 +1,11 @@
|
|||
package cache
|
||||
|
||||
import (
|
||||
"encoding/binary"
|
||||
"hash/fnv"
|
||||
"crypto/rand"
|
||||
"math"
|
||||
"math/big"
|
||||
insecurerand "math/rand"
|
||||
"os"
|
||||
"runtime"
|
||||
"time"
|
||||
)
|
||||
|
@ -10,8 +13,9 @@ import (
|
|||
// This is an experimental and unexported (for now) attempt at making a cache
|
||||
// with better algorithmic complexity than the standard one, namely by
|
||||
// preventing write locks of the entire cache when an item is added. As of the
|
||||
// time of writing, the overhead of selecting buckets is much too great to be
|
||||
// worth it except perhaps with extremely large caches.
|
||||
// time of writing, the overhead of selecting buckets results in cache
|
||||
// operations being about twice as slow as for the standard cache with small
|
||||
// total cache sizes, and faster for larger ones.
|
||||
//
|
||||
// See cache_test.go for a few benchmarks.
|
||||
|
||||
|
@ -20,16 +24,46 @@ type unexportedShardedCache struct {
|
|||
}
|
||||
|
||||
type shardedCache struct {
|
||||
seed uint32
|
||||
m uint32
|
||||
cs []*cache
|
||||
janitor *shardedJanitor
|
||||
}
|
||||
|
||||
// djb2 with better shuffling. 5x faster than FNV with the hash.Hash overhead.
|
||||
func djb33(seed uint32, k string) uint32 {
|
||||
var (
|
||||
l = uint32(len(k))
|
||||
d = 5381 + seed + l
|
||||
i = uint32(0)
|
||||
)
|
||||
// Why is all this 5x faster than a for loop?
|
||||
if l >= 4 {
|
||||
for i < l-4 {
|
||||
d = (d * 33) ^ uint32(k[i])
|
||||
d = (d * 33) ^ uint32(k[i+1])
|
||||
d = (d * 33) ^ uint32(k[i+2])
|
||||
d = (d * 33) ^ uint32(k[i+3])
|
||||
i += 4
|
||||
}
|
||||
}
|
||||
switch l - i {
|
||||
case 1:
|
||||
case 2:
|
||||
d = (d * 33) ^ uint32(k[i])
|
||||
case 3:
|
||||
d = (d * 33) ^ uint32(k[i])
|
||||
d = (d * 33) ^ uint32(k[i+1])
|
||||
case 4:
|
||||
d = (d * 33) ^ uint32(k[i])
|
||||
d = (d * 33) ^ uint32(k[i+1])
|
||||
d = (d * 33) ^ uint32(k[i+2])
|
||||
}
|
||||
return d ^ (d >> 16)
|
||||
}
|
||||
|
||||
func (sc *shardedCache) bucket(k string) *cache {
|
||||
h := fnv.New32()
|
||||
h.Write([]byte(k))
|
||||
n := binary.BigEndian.Uint32(h.Sum(nil))
|
||||
return sc.cs[n%sc.m]
|
||||
return sc.cs[djb33(sc.seed, k)%sc.m]
|
||||
}
|
||||
|
||||
func (sc *shardedCache) Set(k string, x interface{}, d time.Duration) {
|
||||
|
@ -70,6 +104,19 @@ func (sc *shardedCache) DeleteExpired() {
|
|||
}
|
||||
}
|
||||
|
||||
// Returns the items in the cache. This may include items that have expired,
|
||||
// but have not yet been cleaned up. If this is significant, the Expiration
|
||||
// fields of the items should be checked. Note that explicit synchronization
|
||||
// is needed to use a cache and its corresponding Items() return values at
|
||||
// the same time, as the maps are shared.
|
||||
func (sc *shardedCache) Items() []map[string]*Item {
|
||||
res := make([]map[string]*Item, len(sc.cs))
|
||||
for i, v := range sc.cs {
|
||||
res[i] = v.Items()
|
||||
}
|
||||
return res
|
||||
}
|
||||
|
||||
func (sc *shardedCache) Flush() {
|
||||
for _, v := range sc.cs {
|
||||
v.Flush()
|
||||
|
@ -107,8 +154,18 @@ func runShardedJanitor(sc *shardedCache, ci time.Duration) {
|
|||
}
|
||||
|
||||
func newShardedCache(n int, de time.Duration) *shardedCache {
|
||||
max := big.NewInt(0).SetUint64(uint64(math.MaxUint32))
|
||||
rnd, err := rand.Int(rand.Reader, max)
|
||||
var seed uint32
|
||||
if err != nil {
|
||||
os.Stderr.Write([]byte("WARNING: go-cache's newShardedCache failed to read from the system CSPRNG (/dev/urandom or equivalent.) Your system's security may be compromised. Continuing with an insecure seed.\n"))
|
||||
seed = insecurerand.Uint32()
|
||||
} else {
|
||||
seed = uint32(rnd.Uint64())
|
||||
}
|
||||
sc := &shardedCache{
|
||||
m: uint32(n - 1),
|
||||
seed: seed,
|
||||
m: uint32(n),
|
||||
cs: make([]*cache, n),
|
||||
}
|
||||
for i := 0; i < n; i++ {
|
||||
|
@ -121,7 +178,7 @@ func newShardedCache(n int, de time.Duration) *shardedCache {
|
|||
return sc
|
||||
}
|
||||
|
||||
func unexportedNewSharded(shards int, defaultExpiration, cleanupInterval time.Duration) *unexportedShardedCache {
|
||||
func unexportedNewSharded(defaultExpiration, cleanupInterval time.Duration, shards int) *unexportedShardedCache {
|
||||
if defaultExpiration == 0 {
|
||||
defaultExpiration = -1
|
||||
}
|
||||
|
|
|
@ -0,0 +1,68 @@
|
|||
package cache
|
||||
|
||||
import (
|
||||
"strconv"
|
||||
"sync"
|
||||
"testing"
|
||||
)
|
||||
|
||||
// func TestDjb33(t *testing.T) {
|
||||
// }
|
||||
|
||||
var shardedKeys = []string{
|
||||
"f",
|
||||
"fo",
|
||||
"foo",
|
||||
"barf",
|
||||
"barfo",
|
||||
"foobar",
|
||||
"bazbarf",
|
||||
"bazbarfo",
|
||||
"bazbarfoo",
|
||||
"foobarbazq",
|
||||
"foobarbazqu",
|
||||
"foobarbazquu",
|
||||
"foobarbazquux",
|
||||
}
|
||||
|
||||
func TestShardedCache(t *testing.T) {
|
||||
tc := unexportedNewSharded(DefaultExpiration, 0, 13)
|
||||
for _, v := range shardedKeys {
|
||||
tc.Set(v, "value", DefaultExpiration)
|
||||
}
|
||||
}
|
||||
|
||||
func BenchmarkShardedCacheGet(b *testing.B) {
|
||||
b.StopTimer()
|
||||
tc := unexportedNewSharded(DefaultExpiration, 0, 10)
|
||||
tc.Set("foobarba", "zquux", DefaultExpiration)
|
||||
b.StartTimer()
|
||||
for i := 0; i < b.N; i++ {
|
||||
tc.Get("foobarba")
|
||||
}
|
||||
}
|
||||
|
||||
func BenchmarkShardedCacheGetManyConcurrent(b *testing.B) {
|
||||
b.StopTimer()
|
||||
n := 10000
|
||||
tsc := unexportedNewSharded(DefaultExpiration, 0, 20)
|
||||
keys := make([]string, n)
|
||||
for i := 0; i < n; i++ {
|
||||
k := "foo" + strconv.Itoa(n)
|
||||
keys[i] = k
|
||||
tsc.Set(k, "bar", DefaultExpiration)
|
||||
}
|
||||
each := b.N / n
|
||||
wg := new(sync.WaitGroup)
|
||||
wg.Add(n)
|
||||
for _, v := range keys {
|
||||
go func() {
|
||||
for j := 0; j < each; j++ {
|
||||
tsc.Get(v)
|
||||
}
|
||||
wg.Done()
|
||||
}()
|
||||
}
|
||||
b.StartTimer()
|
||||
wg.Wait()
|
||||
}
|
Loading…
Reference in New Issue