Miscellaneous changes during perf test

pkg/disk/batcher/low_latency.go

@@ -170,7 +170,7 @@ func (w *LowLatency[T]) descBatch(logger logr.Logger, t time.Time, requests map[
 	}
 	// Not found
 	sendToAll(requests, getResponse[*T]{})
-	logger.V(3).Info("got batch", "n", len(thisBatch),
+	logger.V(2).Info("got batch", "n", len(thisBatch),
 		"requestID", resp.RequestID, "duration", w.clk.Since(t), "wait", t.Sub(firstTime))
 }
 

pkg/disk/cloud.go

@@ -158,7 +158,7 @@ func (ad *DiskAttachDetach) findDevice(ctx context.Context, diskID, serial strin
 // Returns device path if fromNode, disk serial number otherwise.
 func (ad *DiskAttachDetach) attachDisk(ctx context.Context, diskID, nodeID string, fromNode bool) (string, error) {
 	logger := klog.FromContext(ctx)
-	logger.V(2).Info("Starting Do AttachDisk", "instanceID", nodeID, "region", GlobalConfigVar.Region)
+	logger.V(2).Info("Starting Do AttachDisk")
 
 	ecsClient := GlobalConfigVar.EcsClient
 	// Step 1: check disk status

pkg/disk/disk.go

@@ -246,10 +246,12 @@ func newBatcher(fromNode bool) (waitstatus.StatusWaiter[ecs.Disk], batcher.Batch
 	client := desc.Disk{Client: GlobalConfigVar.EcsClient}
 	ctx := context.Background()
 	interval := 1 * time.Second
+	max := 2 * time.Second
 	if fromNode {
 		interval = 2 * time.Second // We have many nodes, use longer interval to avoid throttling
+		max = 3 * time.Second
 	}
-	waiter := waitstatus.NewBatched(client, clock.RealClock{}, interval, 3*time.Second)
+	waiter := waitstatus.NewBatched(client, clock.RealClock{}, interval, max)
 	go waiter.Run(ctx)
 
 	b := batcher.NewLowLatency(client, clock.RealClock{}, 1*time.Second, 8)

pkg/disk/utils.go

@@ -109,6 +109,7 @@ var ecsOpenAPITransport = http.Transport{
 	ForceAttemptHTTP2:     true,
 	MaxIdleConns:          100,
 	MaxIdleConnsPerHost:   100, // Set this equal to MaxIdleConns as we should only talk to one endpoint with this Transport instance.
+	MaxConnsPerHost:       500, // Protect our backend. Should be large enough to handle any workload.
 	IdleConnTimeout:       90 * time.Second,
 	TLSHandshakeTimeout:   10 * time.Second,
 	ExpectContinueTimeout: 1 * time.Second,

pkg/disk/waitstatus/batched.go

@@ -117,7 +117,7 @@ func (w *Batched[T]) Run(ctx context.Context) {
 		case r := <-w.feedback:
 			next := w.processFeedback(r)
 			w.idQueue = append(w.idQueue, next...)
-			logger.V(4).Info("poll response processed", "queueDepth", len(w.idQueue), "requeue", len(next))
+			logger.V(2).Info("poll response processed", "queueDepth", len(w.idQueue), "requeue", len(next))
 		case t := <-pollChan:
 			logger.V(4).Info("starting poll", "queueDepth", len(w.idQueue))
 			w.idQueue = w.poll(t, w.idQueue)

pkg/features/features.go

@@ -16,15 +16,17 @@ const (
 	DiskADController featuregate.Feature = "DiskADController"
 
 	// Attach multiple disks to the same node in parallel.
-	// ECS don't allow parallel attach/detach to a node by default.
+	// ECS don't allow parallel attach to a node by default.
 	// Enable this if you need faster attach, and only if your UID is whitelisted (by open a ticket),
 	// or you have the supportConcurrencyAttach=true tag on your ECS instance.
 	//
 	// Only effective when DiskADController is also enabled.
 	DiskParallelAttach featuregate.Feature = "DiskParallelAttach"
 
 	// Detach multiple disks from the same node in parallel.
-	// ECS does not allow parallel detach from a node currently. This feature gate is reserved for future use.
+	// ECS does not allow parallel detach from a node by default.
+	// Enable this if you need faster detach, and only if your UID is whitelisted (by open a ticket),
+	// or you have the supportConcurrencyDetach=true tag on your ECS instance.
 	//
 	// Only effective when DiskADController is also enabled.
 	DiskParallelDetach featuregate.Feature = "DiskParallelDetach"