Merge pull request kubernetes#128999 from macsko/improve_goroutines_metric_writes_in_parallelizer_until

Improve Goroutines metric calls in parallelizer.Until

Author: k8s-ci-robot

pkg/scheduler/framework/parallelize/parallelism.go

@@ -51,15 +51,28 @@ func chunkSizeFor(n, parallelism int) int {
 	return s
 }
 
+// numWorkersForChunkSize returns number of workers (goroutines)
+// that will be created in workqueue.ParallelizeUntil
+// for given parallelism, pieces and chunkSize values.
+func numWorkersForChunkSize(parallelism, pieces, chunkSize int) int {
+	chunks := (pieces + chunkSize - 1) / chunkSize
+	if chunks < parallelism {
+		return chunks
+	}
+	return parallelism
+}
+
 // Until is a wrapper around workqueue.ParallelizeUntil to use in scheduling algorithms.
 // A given operation will be a label that is recorded in the goroutine metric.
 func (p Parallelizer) Until(ctx context.Context, pieces int, doWorkPiece workqueue.DoWorkPieceFunc, operation string) {
+	chunkSize := chunkSizeFor(pieces, p.parallelism)
+	workers := numWorkersForChunkSize(p.parallelism, pieces, chunkSize)
+
 	goroutinesMetric := metrics.Goroutines.WithLabelValues(operation)
-	withMetrics := func(piece int) {
-		goroutinesMetric.Inc()
-		doWorkPiece(piece)
-		goroutinesMetric.Dec()
-	}
+	// Calling single Add with workers' count is more efficient than calling Inc or Dec per each work piece.
+	// This approach improves performance of some plugins (affinity, topology spreading) as well as preemption.
+	goroutinesMetric.Add(float64(workers))
+	defer goroutinesMetric.Add(float64(-workers))
 
-	workqueue.ParallelizeUntil(ctx, p.parallelism, pieces, withMetrics, workqueue.WithChunkSize(chunkSizeFor(pieces, p.parallelism)))
+	workqueue.ParallelizeUntil(ctx, p.parallelism, pieces, doWorkPiece, workqueue.WithChunkSize(chunkSize))
 }