OCPBUGS-13747: 4.13: kubelet/cm: disable cpu load balancing on slices when using static cpu manager policy

pkg/kubelet/cm/cgroup_manager_linux.go

@@ -38,7 +38,6 @@ import (
 	utilruntime "k8s.io/apimachinery/pkg/util/runtime"
 	"k8s.io/apimachinery/pkg/util/sets"
 	cmutil "k8s.io/kubernetes/pkg/kubelet/cm/util"
-	"k8s.io/kubernetes/pkg/kubelet/managed"
 	"k8s.io/kubernetes/pkg/kubelet/metrics"
 )
 
@@ -148,6 +147,10 @@ type cgroupManagerImpl struct {
 
 	// useSystemd tells if systemd cgroup manager should be used.
 	useSystemd bool
+
+	// cpuLoadBalanceDisable tells whether kubelet should disable
+	// cpu load balancing on new cgroups it creates.
+	cpuLoadBalanceDisable bool
 }
 
 // Make sure that cgroupManagerImpl implements the CgroupManager interface
@@ -474,10 +477,10 @@ func (m *cgroupManagerImpl) Create(cgroupConfig *CgroupConfig) error {
 		utilruntime.HandleError(fmt.Errorf("cgroup manager.Set failed: %w", err))
 	}
 
-	// Disable cpuset.sched_load_balance for all cgroups Kubelet creates when kubelet has workloads to manage.
+	// Disable cpuset.sched_load_balance for all cgroups Kubelet creates.
 	// This way, CRI can disable sched_load_balance for pods that must have load balance
 	// disabled, but the slices can contain all cpus (as the guaranteed cpus are known dynamically).
-	if managed.IsEnabled() && !libcontainercgroups.IsCgroup2UnifiedMode() {
+	if m.cpuLoadBalanceDisable && !libcontainercgroups.IsCgroup2UnifiedMode() {
 		path := manager.Path("cpuset")
 		if path == "" {
 			return fmt.Errorf("Failed to find cpuset for newly created cgroup")
@@ -572,3 +575,7 @@ func (m *cgroupManagerImpl) MemoryUsage(name CgroupName) (int64, error) {
 	val, err := fscommon.GetCgroupParamUint(path, file)
 	return int64(val), err
 }
+
+func (m *cgroupManagerImpl) SetCPULoadBalanceDisable() {
+	m.cpuLoadBalanceDisable = true
+}

pkg/kubelet/cm/cgroup_manager_unsupported.go

@@ -77,6 +77,9 @@ func (m *unsupportedCgroupManager) ReduceCPULimits(cgroupName CgroupName) error
 	return nil
 }
 
+func (m *unsupportedCgroupManager) SetCPULoadBalanceDisable() {
+}
+
 var RootCgroupName = CgroupName([]string{})
 
 func NewCgroupName(base CgroupName, components ...string) CgroupName {

pkg/kubelet/cm/container_manager_linux.go

@@ -250,6 +250,9 @@ func NewContainerManager(mountUtil mount.Interface, cadvisorInterface cadvisor.I
 	// Turn CgroupRoot from a string (in cgroupfs path format) to internal CgroupName
 	cgroupRoot := ParseCgroupfsToCgroupName(nodeConfig.CgroupRoot)
 	cgroupManager := NewCgroupManager(subsystems, nodeConfig.CgroupDriver)
+	if nodeConfig.CPUManagerPolicy == string(cpumanager.PolicyStatic) {
+		cgroupManager.SetCPULoadBalanceDisable()
+	}
 	// Check if Cgroup-root actually exists on the node
 	if nodeConfig.CgroupsPerQOS {
 		// this does default to / when enabled, but this tests against regressions.

pkg/kubelet/cm/types.go

@@ -84,6 +84,8 @@ type CgroupManager interface {
 	ReduceCPULimits(cgroupName CgroupName) error
 	// MemoryUsage returns current memory usage of the specified cgroup, as read from the cgroupfs.
 	MemoryUsage(name CgroupName) (int64, error)
+	// Toggle whether CPU load balancing should be disabled for new cgroups the kubelet creates
+	SetCPULoadBalanceDisable()
 }
 
 // QOSContainersInfo stores the names of containers per qos