diff --git a/utils/cpuload/cpuload.go b/utils/cpuload/cpuload.go
index 1679a8fd..81b7066d 100644
--- a/utils/cpuload/cpuload.go
+++ b/utils/cpuload/cpuload.go
@@ -37,18 +37,18 @@ type CpuLoadReader interface {
 }
 
 func New() (CpuLoadReader, error) {
-	// First try to create a netlink based load reader.
-	reader, err := netlink.New()
-	if err == nil {
-		glog.Info("Using a netlink-based load reader")
-		return reader, nil
-	}
-	glog.V(1).Infof("failed to create a netlink-based cpu load read: %v", err)
-	// Netlink based load reader doesn't work inside namespaces. Fall back to using scheddebug.
+	// First try to create a scheddebug based load reader.
 	schedReader, schedErr := scheddebug.New()
-	if schedErr != nil {
+	if schedErr == nil {
+		glog.Info("Using a sched debug based load reader")
+		return schedReader, nil
+	}
+	glog.V(1).Infof("failed to create a scheddebug-based cpu load reader: %v", schedErr)
+	// netlink gives us more data than scheddebug, but it doesn't work inside network namespaces. It also needs to be hierarchical.
+	reader, err := netlink.New()
+	if err != nil {
 		return nil, fmt.Errorf("failed to create any cpu load reader - netlink based (%v), scheddebug based (%v)", err, schedErr)
 	}
-	glog.Info("Using a sched debug based load reader")
-	return schedReader, nil
+	glog.Info("Using a netlink-based load reader")
+	return reader, nil
 }