dedicated upload processor

avoid thundering effect of overloading volume servers

weed/operation/upload_processor.go

@@ -0,0 +1,61 @@
+package operation
+
+import (
+	"reflect"
+	"runtime"
+	"sync"
+	"sync/atomic"
+)
+
+type OperationRequest func()
+
+var (
+	requestSlots          = uint32(32)
+	requests              = make([]chan OperationRequest, requestSlots) // increase slots to increase fairness
+	ConcurrentUploadLimit = int32(runtime.NumCPU())                     // directly related to memory usage
+	concurrentLimitCond   = sync.NewCond(new(sync.Mutex))
+	concurrentUpload      int32
+)
+
+func init() {
+
+	for i := 0; i < int(requestSlots); i++ {
+		requests[i] = make(chan OperationRequest)
+	}
+
+	cases := make([]reflect.SelectCase, requestSlots)
+	for i, ch := range requests {
+		cases[i] = reflect.SelectCase{Dir: reflect.SelectRecv, Chan: reflect.ValueOf(ch)}
+	}
+
+	go func() {
+		for {
+			_, value, ok := reflect.Select(cases)
+			if !ok {
+				continue
+			}
+
+			request := value.Interface().(OperationRequest)
+
+			concurrentLimitCond.L.Lock()
+			for atomic.LoadInt32(&concurrentUpload) > ConcurrentUploadLimit {
+				concurrentLimitCond.Wait()
+			}
+			atomic.AddInt32(&concurrentUpload, 1)
+			concurrentLimitCond.L.Unlock()
+
+			go func() {
+				defer atomic.AddInt32(&concurrentUpload, -1)
+				defer concurrentLimitCond.Signal()
+				request()
+			}()
+
+		}
+	}()
+
+}
+
+func AsyncOutOfOrderProcess(slotKey uint32, request OperationRequest) {
+	index := slotKey % requestSlots
+	requests[index] <- request
+}