Merge pull request containers#282 from klihub/devel/topology-aware/co…

…nfigurable-allocation-priority

topology-aware: configurable allocation priority

cmd/plugins/topology-aware/policy/coldstart_test.go

@@ -95,7 +95,8 @@ func TestColdStart(t *testing.T) {
 				allocations: allocations{
 					grants: make(map[string]Grant, 0),
 				},
-				options: &policyapi.BackendOptions{},
+				options:      &policyapi.BackendOptions{},
+				cpuAllocator: &mockCPUAllocator{},
 			}
 			policy.allocations.policy = policy
 			policy.options.SendEvent = sendEvent

cmd/plugins/topology-aware/policy/mocks_test.go

@@ -20,7 +20,9 @@ import (
 
 	nri "github.com/containerd/nri/pkg/api"
 	resmgr "github.com/containers/nri-plugins/pkg/apis/resmgr/v1alpha1"
+	"github.com/containers/nri-plugins/pkg/cpuallocator"
 	"github.com/containers/nri-plugins/pkg/resmgr/cache"
+	"github.com/containers/nri-plugins/pkg/sysfs"
 	system "github.com/containers/nri-plugins/pkg/sysfs"
 	"github.com/containers/nri-plugins/pkg/topology"
 	"github.com/containers/nri-plugins/pkg/utils/cpuset"
@@ -103,6 +105,22 @@ func (p *mockCPUPackage) DieNodeIDs(idset.ID) []idset.ID {
 	return []idset.ID{}
 }
 
+func (p *mockCPUPackage) DieClusterIDs(idset.ID) []idset.ID {
+	return []idset.ID{}
+}
+
+func (p *mockCPUPackage) DieClusterCPUSet(idset.ID, idset.ID) cpuset.CPUSet {
+	return cpuset.New()
+}
+
+func (p *mockCPUPackage) LogicalDieClusterIDs(idset.ID) []idset.ID {
+	return []idset.ID{}
+}
+
+func (p *mockCPUPackage) LogicalDieClusterCPUSet(idset.ID, idset.ID) cpuset.CPUSet {
+	return cpuset.New()
+}
+
 func (p *mockCPUPackage) SstInfo() *sst.SstPackageInfo {
 	return &sst.SstPackageInfo{}
 }
@@ -156,6 +174,33 @@ func (c *mockCPU) SstClos() int {
 	return -1
 }
 
+func (c *mockCPU) CacheCount() int {
+	return 0
+}
+func (c *mockCPU) GetCaches() []*sysfs.Cache {
+	panic("unimplemented")
+}
+func (c *mockCPU) GetCachesByLevel(int) []*sysfs.Cache {
+	panic("unimplemented")
+}
+func (c *mockCPU) GetCacheByIndex(int) *sysfs.Cache {
+	panic("unimplemented")
+}
+func (c *mockCPU) GetLastLevelCaches() []*sysfs.Cache {
+	panic("unimplemented")
+}
+func (c *mockCPU) GetLastLevelCacheCPUSet() cpuset.CPUSet {
+	panic("unimplemented")
+}
+
+func (c *mockCPU) ClusterID() int {
+	return 0
+}
+
+func (c *mockCPU) CoreKind() sysfs.CoreKind {
+	return sysfs.PerformanceCore
+}
+
 type mockSystem struct {
 	isolatedCPU  int
 	nodes        []system.Node
@@ -188,6 +233,27 @@ func (fake *mockSystem) Discover(flags system.DiscoveryFlag) error {
 func (fake *mockSystem) Package(idset.ID) system.CPUPackage {
 	return &mockCPUPackage{}
 }
+func (fake *mockSystem) PossibleCPUs() cpuset.CPUSet {
+	return fake.CPUSet()
+}
+func (fake *mockSystem) PresentCPUs() cpuset.CPUSet {
+	return fake.CPUSet()
+}
+func (fake *mockSystem) OnlineCPUs() cpuset.CPUSet {
+	return fake.CPUSet()
+}
+func (fake *mockSystem) IsolatedCPUs() cpuset.CPUSet {
+	return fake.Isolated()
+}
+func (fake *mockSystem) OfflineCPUs() cpuset.CPUSet {
+	return cpuset.New()
+}
+func (fake *mockSystem) CoreKindCPUs(sysfs.CoreKind) cpuset.CPUSet {
+	return cpuset.New()
+}
+func (fake *mockSystem) AllThreadsForCPUs(cpuset.CPUSet) cpuset.CPUSet {
+	return cpuset.New()
+}
 func (fake *mockSystem) Offlined() cpuset.CPUSet {
 	return cpuset.New()
 }
@@ -629,3 +695,21 @@ func (m *mockCache) OpenFile(string, string, os.FileMode) (*os.File, error) {
 func (m *mockCache) WriteFile(string, string, os.FileMode, []byte) error {
 	panic("unimplemented")
 }
+
+type mockCPUAllocator struct{}
+
+func (m *mockCPUAllocator) AllocateCpus(from *cpuset.CPUSet, cnt int, prefer cpuallocator.CPUPriority) (cpuset.CPUSet, error) {
+	return cpuset.New(0), nil
+}
+
+func (m *mockCPUAllocator) ReleaseCpus(from *cpuset.CPUSet, cnt int, prefer cpuallocator.CPUPriority) (cpuset.CPUSet, error) {
+	return cpuset.New(0), nil
+}
+
+func (m *mockCPUAllocator) GetCPUPriorities() map[cpuallocator.CPUPriority]cpuset.CPUSet {
+	return map[cpuallocator.CPUPriority]cpuset.CPUSet{}
+}
+
+var (
+	_ cpuallocator.CPUAllocator = &mockCPUAllocator{}
+)

cmd/plugins/topology-aware/policy/pod-preferences.go

@@ -42,6 +42,8 @@ const (
 	keyColdStartPreference = "cold-start"
 	// annotation key for reserved pools
 	keyReservedCPUsPreference = "prefer-reserved-cpus"
+	// annotation key for CPU Priority preference
+	keyCpuPriorityPreference = "prefer-cpu-priority"
 
 	// effective annotation key for isolated CPU preference
 	preferIsolatedCPUsKey = keyIsolationPreference + "." + kubernetes.ResmgrKeyNamespace
@@ -53,6 +55,8 @@ const (
 	preferColdStartKey = keyColdStartPreference + "." + kubernetes.ResmgrKeyNamespace
 	// annotation key for reserved pools
 	preferReservedCPUsKey = keyReservedCPUsPreference + "." + kubernetes.ResmgrKeyNamespace
+	// effective annotation key for CPU priority preference
+	preferCpuPriorityKey = keyCpuPriorityPreference + "." + kubernetes.ResmgrKeyNamespace
 )
 
 // cpuClass is a type of CPU to allocate
@@ -153,6 +157,36 @@ func sharedCPUsPreference(pod cache.Pod, container cache.Container) (bool, bool)
 	return preference, true
 }
 
+// cpuPrioPreference returns the CPU priority preference for the given container
+// and whether the container was explicitly annotated with this setting.
+func cpuPrioPreference(pod cache.Pod, container cache.Container, fallback cpuPrio) cpuPrio {
+	key := preferCpuPriorityKey
+	value, ok := pod.GetEffectiveAnnotation(key, container.GetName())
+
+	if !ok {
+		prio := fallback
+		log.Debug("%s: implicit CPU priority preference %q", container.PrettyName(), prio)
+		return prio
+	}
+
+	if value == "default" {
+		prio := defaultPrio
+		log.Debug("%s: explicit default CPU priority preference %q", container.PrettyName(), prio)
+		return prio
+	}
+
+	prio, ok := cpuPrioByName[value]
+	if !ok {
+		log.Error("%s: invalid CPU priority preference %q", container.PrettyName(), value)
+		prio := fallback
+		log.Debug("%s: implicit CPU priority preference %q", container.PrettyName(), prio)
+		return prio
+	}
+
+	log.Debug("%s: explicit CPU priority preference %q", container.PrettyName(), prio)
+	return prio
+}
+
 // memoryTypePreference returns what type of memory should be allocated for the container.
 //
 // If the effective annotations are not found, this function falls back to
@@ -370,7 +404,8 @@ func checkReservedCPUsAnnotations(c cache.Container) (bool, bool) {
 // 2. fraction: amount of fractional CPU in milli-CPU
 // 3. isolate: (bool) whether to prefer isolated full CPUs
 // 4. cpuType: (cpuClass) class of CPU to allocate (reserved vs. normal)
-func cpuAllocationPreferences(pod cache.Pod, container cache.Container) (int, int, bool, cpuClass) {
+// 5. cpuPrio: preferred CPU allocator priority for CPU allocation.
+func cpuAllocationPreferences(pod cache.Pod, container cache.Container) (int, int, bool, cpuClass, cpuPrio) {
 	//
 	// CPU allocation preferences for a container consist of
 	//
@@ -439,60 +474,62 @@ func cpuAllocationPreferences(pod cache.Pod, container cache.Container) (int, in
 	request := reqs.Requests[corev1.ResourceCPU]
 	qosClass := pod.GetQOSClass()
 	fraction := int(request.MilliValue())
+	prio := defaultPrio // ignored for fractional allocations
 
 	// easy cases: kube-system namespace, Burstable or BestEffort QoS class containers
 	preferReserved, explicitReservation := checkReservedCPUsAnnotations(container)
 	switch {
 	case container.PreserveCpuResources():
-		return 0, fraction, false, cpuPreserve
+		return 0, fraction, false, cpuPreserve, prio
 	case preferReserved == true:
-		return 0, fraction, false, cpuReserved
+		return 0, fraction, false, cpuReserved, prio
 	case checkReservedPoolNamespaces(namespace) && !explicitReservation:
-		return 0, fraction, false, cpuReserved
+		return 0, fraction, false, cpuReserved, prio
 	case qosClass == corev1.PodQOSBurstable:
-		return 0, fraction, false, cpuNormal
+		return 0, fraction, false, cpuNormal, prio
 	case qosClass == corev1.PodQOSBestEffort:
-		return 0, 0, false, cpuNormal
+		return 0, 0, false, cpuNormal, prio
 	}
 
 	// complex case: Guaranteed QoS class containers
 	cores := fraction / 1000
 	fraction = fraction % 1000
 	preferIsolated, explicitIsolated := isolatedCPUsPreference(pod, container)
 	preferShared, explicitShared := sharedCPUsPreference(pod, container)
+	prio = cpuPrioPreference(pod, container, defaultPrio) // ignored for fractional allocations
 
 	switch {
 	// sub-core CPU request
 	case cores == 0:
-		return 0, fraction, false, cpuNormal
+		return 0, fraction, false, cpuNormal, prio
 		// 1 <= CPU request < 2
 	case cores < 2:
 		// fractional allocation, potentially mixed
 		if fraction > 0 {
 			if preferShared {
-				return 0, 1000*cores + fraction, false, cpuNormal
+				return 0, 1000*cores + fraction, false, cpuNormal, prio
 			}
-			return cores, fraction, preferIsolated, cpuNormal
+			return cores, fraction, preferIsolated, cpuNormal, prio
 		}
 		// non-fractional allocation
 		if preferShared && explicitShared {
-			return 0, 1000*cores + fraction, false, cpuNormal
+			return 0, 1000*cores + fraction, false, cpuNormal, prio
 		}
-		return cores, fraction, preferIsolated, cpuNormal
+		return cores, fraction, preferIsolated, cpuNormal, prio
 		// CPU request >= 2
 	default:
 		// fractional allocation, only mixed if explicitly annotated as unshared
 		if fraction > 0 {
 			if !preferShared && explicitShared {
-				return cores, fraction, preferIsolated && explicitIsolated, cpuNormal
+				return cores, fraction, preferIsolated && explicitIsolated, cpuNormal, prio
 			}
-			return 0, 1000*cores + fraction, false, cpuNormal
+			return 0, 1000*cores + fraction, false, cpuNormal, prio
 		}
 		// non-fractional allocation
 		if preferShared && explicitShared {
-			return 0, 1000 * cores, false, cpuNormal
+			return 0, 1000 * cores, false, cpuNormal, prio
 		}
-		return cores, fraction, preferIsolated && explicitIsolated, cpuNormal
+		return cores, fraction, preferIsolated && explicitIsolated, cpuNormal, prio
 	}
 }
 

cmd/plugins/topology-aware/policy/pod-preferences_test.go

@@ -1038,7 +1038,7 @@ func TestCpuAllocationPreferences(t *testing.T) {
 			}
 			opt.PreferIsolated, opt.PreferShared = tc.preferIsolated, tc.preferShared
 			opt.ReservedPoolNamespaces = tc.reservedPoolNamespaces
-			full, fraction, isolate, cpuType := cpuAllocationPreferences(tc.pod, tc.container)
+			full, fraction, isolate, cpuType, _ := cpuAllocationPreferences(tc.pod, tc.container)
 			if full != tc.expectedFull || fraction != tc.expectedFraction ||
 				isolate != tc.expectedIsolate || cpuType != tc.expectedCpuType {
 				t.Errorf("Expected (%v, %v, %v, %s), but got (%v, %v, %v, %s)",