Add cache with multiple shards to decrease lock contention

pkg/tools/etcd_helper.go

@@ -25,11 +25,11 @@ import (
 	"path"
 	"reflect"
 	"strings"
-	"sync"
 	"time"
 
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/conversion"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/runtime"
+	"github.com/GoogleCloudPlatform/kubernetes/pkg/util"
 	"github.com/coreos/go-etcd/etcd"
 	"github.com/prometheus/client_golang/prometheus"
 
@@ -56,6 +56,18 @@ var (
 				"because two concurrent threads can miss the cache and generate the same entry twice.",
 		},
 	)
+	cacheGetLatency = prometheus.NewSummary(
+		prometheus.SummaryOpts{
+			Name: "etcd_request_cache_get_latencies_summary",
+			Help: "Latency in microseconds of getting an object from etcd cache",
+		},
+	)
+	cacheAddLatency = prometheus.NewSummary(
+		prometheus.SummaryOpts{
+			Name: "etcd_request_cache_add_latencies_summary",
+			Help: "Latency in microseconds of adding an object to etcd cache",
+		},
+	)
 	etcdRequestLatenciesSummary = prometheus.NewSummaryVec(
 		prometheus.SummaryOpts{
 			Name: "etcd_request_latencies_summary",
@@ -65,21 +77,25 @@ var (
 	)
 )
 
-func getTypeName(obj interface{}) string {
-	return reflect.TypeOf(obj).String()
-}
-
-func recordEtcdRequestLatency(verb, resource string, startTime time.Time) {
-	etcdRequestLatenciesSummary.WithLabelValues(verb, resource).Observe(float64(time.Since(startTime) / time.Microsecond))
-}
+const maxEtcdCacheEntries int = 50000
 
 func init() {
 	prometheus.MustRegister(cacheHitCounter)
 	prometheus.MustRegister(cacheMissCounter)
 	prometheus.MustRegister(cacheEntryCounter)
+	prometheus.MustRegister(cacheAddLatency)
+	prometheus.MustRegister(cacheGetLatency)
 	prometheus.MustRegister(etcdRequestLatenciesSummary)
 }
 
+func getTypeName(obj interface{}) string {
+	return reflect.TypeOf(obj).String()
+}
+
+func recordEtcdRequestLatency(verb, resource string, startTime time.Time) {
+	etcdRequestLatenciesSummary.WithLabelValues(verb, resource).Observe(float64(time.Since(startTime) / time.Microsecond))
+}
+
 // EtcdHelper offers common object marshalling/unmarshalling operations on an etcd client.
 type EtcdHelper struct {
 	Client EtcdGetSet
@@ -96,8 +112,7 @@ type EtcdHelper struct {
 	// support multi-object transaction that will result in many objects with the same index.
 	// Number of entries stored in the cache is controlled by maxEtcdCacheEntries constant.
 	// TODO: Measure how much this cache helps after the conversion code is optimized.
-	cache map[uint64]runtime.Object
-	mutex *sync.RWMutex
+	cache util.Cache
 }
 
 // NewEtcdHelper creates a helper that works against objects that use the internal
@@ -108,8 +123,7 @@ func NewEtcdHelper(client EtcdGetSet, codec runtime.Codec, prefix string) EtcdHe
 		Codec:      codec,
 		Versioner:  APIObjectVersioner{},
 		PathPrefix: prefix,
-		cache:      make(map[uint64]runtime.Object),
-		mutex:      new(sync.RWMutex),
+		cache:      util.NewCache(maxEtcdCacheEntries),
 	}
 }
 
@@ -207,16 +221,13 @@ type etcdCache interface {
 	addToCache(index uint64, obj runtime.Object)
 }
 
-const maxEtcdCacheEntries int = 50000
-
 func (h *EtcdHelper) getFromCache(index uint64) (runtime.Object, bool) {
-	var obj runtime.Object
-	func() {
-		h.mutex.RLock()
-		defer h.mutex.RUnlock()
-		obj = h.cache[index]
+	startTime := time.Now()
+	defer func() {
+		cacheGetLatency.Observe(float64(time.Since(startTime) / time.Microsecond))
 	}()
-	if obj != nil {
+	obj, found := h.cache.Get(index)
+	if found {
 		// We should not return the object itself to avoid poluting the cache if someone
 		// modifies returned values.
 		objCopy, err := conversion.DeepCopy(obj)
@@ -232,24 +243,19 @@ func (h *EtcdHelper) getFromCache(index uint64) (runtime.Object, bool) {
 }
 
 func (h *EtcdHelper) addToCache(index uint64, obj runtime.Object) {
+	startTime := time.Now()
+	defer func() {
+		cacheAddLatency.Observe(float64(time.Since(startTime) / time.Microsecond))
+	}()
 	objCopy, err := conversion.DeepCopy(obj)
 	if err != nil {
 		glog.Errorf("Error during DeepCopy of cached object: %q", err)
 		return
 	}
-	h.mutex.Lock()
-	defer h.mutex.Unlock()
-	if _, found := h.cache[index]; !found {
+	isOverwrite := h.cache.Add(index, objCopy)
+	if !isOverwrite {
 		cacheEntryCounter.Inc()
 	}
-	h.cache[index] = objCopy.(runtime.Object)
-	if len(h.cache) > maxEtcdCacheEntries {
-		var randomKey uint64
-		for randomKey = range h.cache {
-			break
-		}
-		delete(h.cache, randomKey)
-	}
 }
 
 // ExtractToList works on a *List api object (an object that satisfies the runtime.IsList

pkg/util/cache.go

@@ -0,0 +1,64 @@
+package util
+
+import (
+	"sync"
+)
+
+const (
+	shardsCount int = 32
+)
+
+type Cache []*cacheShard
+
+func NewCache(maxSize int) Cache {
+	cache := make(Cache, shardsCount)
+	for i := 0; i < shardsCount; i++ {
+		cache[i] = &cacheShard{
+			items:   make(map[uint64]interface{}),
+			maxSize: maxSize / shardsCount,
+		}
+	}
+	return cache
+}
+
+func (c Cache) getShard(index uint64) *cacheShard {
+	return c[index%uint64(shardsCount)]
+}
+
+// Returns true if object already existed, false otherwise.
+func (c *Cache) Add(index uint64, obj interface{}) bool {
+	return c.getShard(index).add(index, obj)
+}
+
+func (c *Cache) Get(index uint64) (obj interface{}, found bool) {
+	return c.getShard(index).get(index)
+}
+
+type cacheShard struct {
+	items map[uint64]interface{}
+	sync.RWMutex
+	maxSize int
+}
+
+// Returns true if object already existed, false otherwise.
+func (s *cacheShard) add(index uint64, obj interface{}) bool {
+	s.Lock()
+	defer s.Unlock()
+	_, isOverwrite := s.items[index]
+	s.items[index] = obj
+	if len(s.items) > s.maxSize {
+		var randomKey uint64
+		for randomKey = range s.items {
+			break
+		}
+		delete(s.items, randomKey)
+	}
+	return isOverwrite
+}
+
+func (s *cacheShard) get(index uint64) (obj interface{}, found bool) {
+	s.RLock()
+	defer s.RUnlock()
+	obj, found = s.items[index]
+	return
+}