Add taints on health events #983
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -24,21 +24,93 @@ import ( | |
| "sync" | ||
|
|
||
| "github.com/NVIDIA/go-nvml/pkg/nvml" | ||
| resourceapi "k8s.io/api/resource/v1" | ||
| "k8s.io/klog/v2" | ||
| ) | ||
|
|
||
| const ( | ||
| FullGPUInstanceID uint32 = 0xFFFFFFFF | ||
| ) | ||
|
|
||
| const ( | ||
| TaintKeyXID = DriverName + "/xid" | ||
| TaintKeyGPULost = DriverName + "/gpu-lost" | ||
| TaintKeyUnmonitored = DriverName + "/unmonitored" | ||
| ) | ||
|
|
||
| // TODO(issue #1001):Remove this hardcoded constant and switch to the upstream | ||
| // resourceapi.DeviceTaintEffectNone once the k8s.io/api dependency is bumped | ||
| // to a version that includes it (KEP-5055 beta). | ||
| // DeviceTaintEffectNone is an informational effect that does not affect | ||
|
Contributor
There was a problem hiding this comment. This hardcoded constant is fragile. If the upstream Consider pinning the vendored
Contributor
Author
There was a problem hiding this comment. mmm its going to be the same https://pkg.go.dev/k8s.io/kubernetes/pkg/apis/resource#DeviceTaintEffect I created the issue: #1001 |
||
| // scheduling or eviction. | ||
| const DeviceTaintEffectNone resourceapi.DeviceTaintEffect = "None" | ||
|
|
||
| // DeviceHealthEventType classifies the category of health event detected by | ||
| // the NVML health monitor. | ||
| type DeviceHealthEventType string | ||
|
|
||
| const ( | ||
| HealthEventXID DeviceHealthEventType = "xid" | ||
| HealthEventGPULost DeviceHealthEventType = "gpu-lost" | ||
| HealthEventUnmonitored DeviceHealthEventType = "unmonitored" | ||
| ) | ||
|
|
||
| // DeviceHealthEvent carries a typed health notification from the NVML health | ||
| // monitor to the driver's event handler, enabling the driver to set the | ||
| // appropriate DRA device taint per the Option A schema (KEP-5055). | ||
| // Devices is a batch: for GPU_LOST and unmonitored events where all affected devices | ||
| // are aggregated into a single event so the consumer applies one ResourceSlice | ||
| // update instead of N. | ||
| type DeviceHealthEvent struct { | ||
| Devices []*AllocatableDevice | ||
| EventType DeviceHealthEventType | ||
| // inspired by NVML Event type and only meaningful for xid errors. | ||
| // may have to create a custom type based on future device-api | ||
| EventData uint64 | ||
| } | ||
|
|
||
| // healthEventToTaint maps a DeviceHealthEvent to the corresponding DRA | ||
| // DeviceTaint using the Option A taint key schema: one key per health | ||
| // dimension under the gpu.nvidia.com domain. | ||
| func healthEventToTaint(monitor deviceHealthMonitor, event *DeviceHealthEvent) *resourceapi.DeviceTaint { | ||
| switch event.EventType { | ||
| case HealthEventXID: | ||
| effect := resourceapi.DeviceTaintEffectNoSchedule | ||
| if monitor != nil && monitor.IsEventNonFatal(event) { | ||
| effect = DeviceTaintEffectNone | ||
| } | ||
| return &resourceapi.DeviceTaint{ | ||
| Key: TaintKeyXID, | ||
| Value: strconv.FormatUint(event.EventData, 10), | ||
| Effect: effect, | ||
| } | ||
| case HealthEventGPULost: | ||
| return &resourceapi.DeviceTaint{ | ||
| Key: TaintKeyGPULost, | ||
| Effect: resourceapi.DeviceTaintEffectNoSchedule, | ||
| } | ||
| case HealthEventUnmonitored: | ||
| return &resourceapi.DeviceTaint{ | ||
| Key: TaintKeyUnmonitored, | ||
| Effect: DeviceTaintEffectNone, | ||
| } | ||
| default: | ||
| klog.Errorf("Unknown health event type %q, defaulting to unmonitored taint", event.EventType) | ||
| return &resourceapi.DeviceTaint{ | ||
| Key: TaintKeyUnmonitored, | ||
| Effect: DeviceTaintEffectNone, | ||
| } | ||
| } | ||
| } | ||
|
|
||
| // For a MIG device the placement is defined by the 3-tuple <parent UUID, GI, CI>. | ||
| // For a full device the returned 3-tuple is the device's uuid and (FullGPUInstanceID) 0xFFFFFFFF for the other two elements. | ||
| type devicePlacementMap map[string]map[uint32]map[uint32]*AllocatableDevice | ||
|
|
||
| type nvmlDeviceHealthMonitor struct { | ||
| nvmllib nvml.Interface | ||
| eventSet nvml.EventSet | ||
| unhealthy chan *DeviceHealthEvent | ||
| deviceByPlacement devicePlacementMap | ||
| skippedXids map[uint64]bool | ||
| wg sync.WaitGroup | ||
|
|
@@ -61,7 +133,7 @@ func newNvmlDeviceHealthMonitor(config *Config, perGPUAllocatable *PerGPUAllocat | |
| all := perGPUAllocatable.GetAllDevices() | ||
| m := &nvmlDeviceHealthMonitor{ | ||
| nvmllib: nvdevlib.nvmllib, | ||
| unhealthy: make(chan *DeviceHealthEvent, len(all)), | ||
| deviceByPlacement: getDevicePlacementMap(all), | ||
| skippedXids: xidsToSkip(config.flags.additionalXidsToIgnore), | ||
| } | ||
|
|
@@ -106,25 +178,25 @@ func (m *nvmlDeviceHealthMonitor) registerEventsForDevices() { | |
| for parentUUID, giMap := range m.deviceByPlacement { | ||
| gpu, ret := m.nvmllib.DeviceGetHandleByUUID(parentUUID) | ||
| if ret != nvml.SUCCESS { | ||
| klog.Warningf("Unable to get device handle from UUID[%s]: %v; marking devices as unmonitored", parentUUID, ret) | ||
| m.sendHealthEventForDevices(giMap, HealthEventUnmonitored) | ||
| continue | ||
| } | ||
|
|
||
| supportedEvents, ret := gpu.GetSupportedEventTypes() | ||
| if ret != nvml.SUCCESS { | ||
| klog.Warningf("unable to determine the supported events for %s: %v; marking devices as unmonitored", parentUUID, ret) | ||
| m.sendHealthEventForDevices(giMap, HealthEventUnmonitored) | ||
| continue | ||
| } | ||
|
|
||
| ret = gpu.RegisterEvents(eventMask&supportedEvents, m.eventSet) | ||
| if ret == nvml.ERROR_NOT_SUPPORTED { | ||
| klog.Warningf("Device %v is too old to support healthchecking.", parentUUID) | ||
| m.sendHealthEventForDevices(giMap, HealthEventUnmonitored) | ||
| } else if ret != nvml.SUCCESS { | ||
| klog.Warningf("unable to register events for %s: %v; marking devices as unmonitored", parentUUID, ret) | ||
| m.sendHealthEventForDevices(giMap, HealthEventUnmonitored) | ||
| } | ||
| } | ||
| } | ||
|
|
@@ -162,8 +234,8 @@ func (m *nvmlDeviceHealthMonitor) run(ctx context.Context) { | |
| // Ref doc: [https://docs.nvidia.com/deploy/nvml-api/group__nvmlEvents.html#group__nvmlEvents_1g9714b0ca9a34c7a7780f87fee16b205c]. | ||
| if ret != nvml.SUCCESS { | ||
| if ret == nvml.ERROR_GPU_IS_LOST { | ||
| klog.Warningf("GPU is lost error: %v; Tainting all devices with %s", ret, TaintKeyGPULost) | ||
| m.sendHealthEventForAllDevices(HealthEventGPULost) | ||
| continue | ||
| } | ||
| klog.V(6).Infof("Error waiting for NVML event: %v. Retrying...", ret) | ||
|
|
@@ -180,19 +252,14 @@ func (m *nvmlDeviceHealthMonitor) run(ctx context.Context) { | |
| continue | ||
| } | ||
|
|
||
| klog.V(4).Infof("Processing event XID=%d event", xid) | ||
| // this seems an extreme action. | ||
| // should we just log the error and proceed anyway. | ||
| // TODO: look into how to properly handle this error. | ||
| eventUUID, ret := event.Device.GetUUID() | ||
| if ret != nvml.SUCCESS { | ||
| klog.Warningf("Failed to determine uuid for event %v: %v; Tainting all devices with %s", event, ret, TaintKeyGPULost) | ||
| m.sendHealthEventForAllDevices(HealthEventGPULost) | ||
| continue | ||
| } | ||
| affectedDevice := m.deviceByPlacement.get(eventUUID, gi, ci) | ||
|
|
@@ -201,41 +268,65 @@ func (m *nvmlDeviceHealthMonitor) run(ctx context.Context) { | |
| continue | ||
| } | ||
|
|
||
| klog.V(4).Infof("Sending XID=%d health event for device %s", xid, affectedDevice.UUID()) | ||
| m.unhealthy <- &DeviceHealthEvent{ | ||
| Devices: []*AllocatableDevice{affectedDevice}, | ||
| EventType: HealthEventXID, | ||
| EventData: xid, | ||
| } | ||
| } | ||
| } | ||
| } | ||
|
|
||
| func (m *nvmlDeviceHealthMonitor) Unhealthy() <-chan *DeviceHealthEvent { | ||
| return m.unhealthy | ||
| } | ||
|
|
||
| // sendHealthEventForAllDevices aggregates every device across all GPUs into a | ||
| // single batched DeviceHealthEvent so the consumer makes one ResourceSlice | ||
| // update. | ||
| func (m *nvmlDeviceHealthMonitor) sendHealthEventForAllDevices(eventType DeviceHealthEventType) { | ||
| var devices []*AllocatableDevice | ||
| for _, giMap := range m.deviceByPlacement { | ||
| devices = append(devices, flattenMIGDeviceMap(giMap)...) | ||
| } | ||
| m.sendBatchedHealthEvent(devices, eventType) | ||
| } | ||
|
|
||
| // sendHealthEventForDevices aggregates all devices under a single parent GPU | ||
| // into one batched DeviceHealthEvent. | ||
| func (m *nvmlDeviceHealthMonitor) sendHealthEventForDevices(giMap map[uint32]map[uint32]*AllocatableDevice, eventType DeviceHealthEventType) { | ||
| m.sendBatchedHealthEvent(flattenMIGDeviceMap(giMap), eventType) | ||
| } | ||
|
|
||
| // flattenMIGDeviceMap flattens a GI→CI device map into a slice. | ||
| func flattenMIGDeviceMap(giMap map[uint32]map[uint32]*AllocatableDevice) []*AllocatableDevice { | ||
| var devices []*AllocatableDevice | ||
| for _, ciMap := range giMap { | ||
| for _, dev := range ciMap { | ||
| devices = append(devices, dev) | ||
| } | ||
| } | ||
| return devices | ||
| } | ||
|
|
||
| // sendBatchedHealthEvent sends a single DeviceHealthEvent containing all | ||
| // affected devices. Uses a non-blocking send to protect the monitor goroutine | ||
| // from deadlocks when the channel is full. | ||
| func (m *nvmlDeviceHealthMonitor) sendBatchedHealthEvent(devices []*AllocatableDevice, eventType DeviceHealthEventType) { | ||
| if len(devices) == 0 { | ||
| return | ||
| } | ||
| event := &DeviceHealthEvent{ | ||
| Devices: devices, | ||
| EventType: eventType, | ||
| } | ||
| select { | ||
| case m.unhealthy <- event: | ||
| klog.V(6).Infof("Sent batched %s health event for %d device(s)", eventType, len(devices)) | ||
| default: | ||
| klog.Errorf("Health event channel full; dropping batched %s event for %d device(s)", eventType, len(devices)) | ||
| } | ||
| } | ||
|
|
||
| // The purpose of this function is to allow for a O(1) lookup of | ||
|
|
@@ -268,7 +359,7 @@ func getDevicePlacementMap(allocatable AllocatableDevices) devicePlacementMap { | |
| continue | ||
| } | ||
| giID = d.MigStatic.gIInfo.Id | ||
| ciID = d.MigStatic.cIInfo.Id | ||
|
|
||
| default: | ||
| // This may be a problem; and should be logged | ||
|
|
@@ -304,8 +395,8 @@ func (p devicePlacementMap) get(uuid string, gi, ci uint32) *AllocatableDevice { | |
| } | ||
|
|
||
| // getAdditionalXids returns a list of additional Xids to skip from the specified string. | ||
| // The input is treated as a comma-separated string and all valid uint64 values are considered as Xid values. | ||
| // Invalid values are ignored. | ||
| // TODO: add list of EXPLICIT XIDs from [https://github.com/NVIDIA/k8s-device-plugin/pull/1443]. | ||
| func getAdditionalXids(input string) []uint64 { | ||
| if input == "" { | ||
|
|
@@ -353,3 +444,13 @@ func xidsToSkip(additionalXids string) map[uint64]bool { | |
| } | ||
| return skippedXids | ||
| } | ||
|
|
||
| // IsEventNonFatal evaluates whether a hardware event is considered an application-level | ||
| // warning (None) rather than a critical hardware failure (NoSchedule). | ||
| // Currently, it only checks for XID events. | ||
| func (m *nvmlDeviceHealthMonitor) IsEventNonFatal(event *DeviceHealthEvent) bool { | ||
| if event.EventType == HealthEventXID { | ||
| return m.skippedXids[event.EventData] | ||
| } | ||
| return false | ||
| } | ||
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
AddOrUpdateTaintdeduplicates by key, so if a device gets XID 48 then XID 63, the value overwrites to"63"and the first XID info is lost. This is consistent with Option A's one-key-per-dimension design, but should be documented — e.g., a comment here noting that only the most recent XID is retained per device.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
done