fix: rename compute-isolation to isolation, add integralDecayFactor for PID controller, bump helm version (#418)

* fix: rename compute-isolation to isolation, add integralDecayFactor for PID controller, bump helm version

* fix: bump dependency versions

Author: Code2Life

api/v1/schedulingconfigtemplate_types.go

@@ -297,6 +297,9 @@ type ElasticRateLimitParameters struct {
 	// token bucket min and max
 	CapacityMin string `json:"capacityMin,omitempty"`
 	CapacityMax string `json:"capacityMax,omitempty"`
+
+	// Decay factor for integral term in PID controller, to avoid integral windup
+	IntegralDecayFactor string `json:"integralDecayFactor,omitempty"`
 }
 
 // SchedulingConfigTemplateStatus defines the observed state of SchedulingConfigTemplate.

api/v1/workloadprofile_types.go

@@ -60,8 +60,8 @@ type WorkloadProfileSpec struct {
 
 	// +optional
 	// +kubebuilder:default=soft
-	// How to isolate computing resources, could be `shared` or `soft` or `hard`
-	ComputeIsolation ComputingIsolationMode `json:"computeIsolation,omitempty"`
+	// How to isolate resources, could be `shared` or `soft` or `hard` or `partitioned`
+	Isolation IsolationModeType `json:"isolation,omitempty"`
 
 	// +optional
 	// GPUModel specifies the required GPU model (e.g., "A100", "H100")
@@ -91,13 +91,30 @@ type WorkloadProfileSpec struct {
 	WorkerPodTemplate *runtime.RawExtension `json:"workerPodTemplate,omitempty"`
 }
 
-// +kubebuilder:validation:Enum=shared;soft;hard
-type ComputingIsolationMode string
+// +kubebuilder:validation:Enum=shared;soft;hard;partitioned
+type IsolationModeType string
 
 const (
-	ComputingIsolationModeShared = "shared"
-	ComputingIsolationModeSoft   = "soft"
-	ComputingIsolationModeHard   = "hard"
+	// no limits, rely on GPU built-in time-slicing, each process gets equal share of GPU
+	// Pros: simple and stable, no performance overhead, maximize GPU utilization when well-scheduled
+	// Cons: can not auto-scale and differentiate QoS levels, TFLOPs limit does not take effect, may cause resource contention
+	IsolationModeShared = "shared"
+
+	// default isolation mode, use Proportional-Integral-Derivative controller to isolate computing resources and assign time slices
+	// Pros: can set QoS levels for different workloads, TFLOPs limit is relatively accurate
+	// Cons: ~1% performance overhead, resource contention may occur when burst credits are consumed
+	IsolationModeSoft = "soft"
+
+	// use dedicated SMs to isolate computing resources
+	// Pros: better performance isolation, no performance overhead, oversubscription is possible
+	// Cons: can not auto-scale dynamically, percent may not 1%/1TFLOPs accuracy, coupled with GPU vendor's SM partitioning implementation
+	// NOTE: this can only be used in Remote or Local+SidecarWorker mode, not supported in LocalGPU mode (because no TensorFusion Worker)
+	IsolationModeHard = "hard"
+
+	// use GPU driver level partitioning to isolate resources, need hardware support
+	// Pros: no performance overhead, no resource contention, fully-isolated
+	// Cons: not supported by all GPUs/XPUs, oversubscription is not possible
+	IsolationModePartitioned = "partitioned"
 )
 
 func (t WorkloadProfileSpec) IsDynamicReplica() bool {

charts/tensor-fusion/Chart.yaml

@@ -15,10 +15,10 @@ type: application
 # This is the chart version. This version number should be incremented each time you make changes
 # to the chart and its templates, including the app version.
 # Versions are expected to follow Semantic Versioning (https://semver.org/)
-version: 1.7.2
+version: 1.7.3
 
 # This is the version number of the application being deployed. This version number should be
 # incremented each time you make changes to the application. Versions are not expected to
 # follow Semantic Versioning. They should reflect the version the application is using.
 # It is recommended to use it with quotes.
-appVersion: "1.48.3"
+appVersion: "1.48.5"

charts/tensor-fusion/crds/tensor-fusion.ai_schedulingconfigtemplates.yaml

@@ -274,6 +274,10 @@ spec:
                         description: Filter ineffective requests from rate limit,
                           0.0 to 1.0
                         type: string
+                      integralDecayFactor:
+                        description: Decay factor for integral term in PID controller,
+                          to avoid integral windup
+                        type: string
                       kd:
                         type: string
                       ki:

charts/tensor-fusion/crds/tensor-fusion.ai_tensorfusionworkloads.yaml

@@ -238,15 +238,6 @@ spec:
                       type: object
                     type: array
                 type: object
-              computeIsolation:
-                default: soft
-                description: How to isolate computing resources, could be `shared`
-                  or `soft` or `hard`
-                enum:
-                - shared
-                - soft
-                - hard
-                type: string
               gpuCount:
                 description: The number of GPUs to be used by the workload, default
                   to 1
@@ -266,6 +257,16 @@ spec:
                 description: Schedule the workload to the same GPU server that runs
                   vGPU worker for best performance, default to false
                 type: boolean
+              isolation:
+                default: soft
+                description: How to isolate resources, could be `shared` or `soft`
+                  or `hard` or `partitioned`
+                enum:
+                - shared
+                - soft
+                - hard
+                - partitioned
+                type: string
               nodeAffinity:
                 description: NodeAffinity specifies the node affinity requirements
                   for the workload

charts/tensor-fusion/crds/tensor-fusion.ai_workloadprofiles.yaml

@@ -225,15 +225,6 @@ spec:
                       type: object
                     type: array
                 type: object
-              computeIsolation:
-                default: soft
-                description: How to isolate computing resources, could be `shared`
-                  or `soft` or `hard`
-                enum:
-                - shared
-                - soft
-                - hard
-                type: string
               gpuCount:
                 description: The number of GPUs to be used by the workload, default
                   to 1
@@ -253,6 +244,16 @@ spec:
                 description: Schedule the workload to the same GPU server that runs
                   vGPU worker for best performance, default to false
                 type: boolean
+              isolation:
+                default: soft
+                description: How to isolate resources, could be `shared` or `soft`
+                  or `hard` or `partitioned`
+                enum:
+                - shared
+                - soft
+                - hard
+                - partitioned
+                type: string
               nodeAffinity:
                 description: NodeAffinity specifies the node affinity requirements
                   for the workload

charts/tensor-fusion/values.yaml

@@ -31,7 +31,7 @@ controller:
   image:
     repository: tensorfusion/tensor-fusion-operator
     # Overrides the image tag whose default is the chart appVersion.
-    tag: "1.48.2"
+    tag: "1.48.5"
   # This is for setting Kubernetes Annotations to a Pod.
   # For more information checkout: https://kubernetes.io/docs/concepts/overview/working-with-objects/annotations/ 
 

config/crd/bases/tensor-fusion.ai_schedulingconfigtemplates.yaml

@@ -274,6 +274,10 @@ spec:
                         description: Filter ineffective requests from rate limit,
                           0.0 to 1.0
                         type: string
+                      integralDecayFactor:
+                        description: Decay factor for integral term in PID controller,
+                          to avoid integral windup
+                        type: string
                       kd:
                         type: string
                       ki:

config/crd/bases/tensor-fusion.ai_tensorfusionworkloads.yaml

@@ -238,15 +238,6 @@ spec:
                       type: object
                     type: array
                 type: object
-              computeIsolation:
-                default: soft
-                description: How to isolate computing resources, could be `shared`
-                  or `soft` or `hard`
-                enum:
-                - shared
-                - soft
-                - hard
-                type: string
               gpuCount:
                 description: The number of GPUs to be used by the workload, default
                   to 1
@@ -266,6 +257,16 @@ spec:
                 description: Schedule the workload to the same GPU server that runs
                   vGPU worker for best performance, default to false
                 type: boolean
+              isolation:
+                default: soft
+                description: How to isolate resources, could be `shared` or `soft`
+                  or `hard` or `partitioned`
+                enum:
+                - shared
+                - soft
+                - hard
+                - partitioned
+                type: string
               nodeAffinity:
                 description: NodeAffinity specifies the node affinity requirements
                   for the workload

config/crd/bases/tensor-fusion.ai_workloadprofiles.yaml

@@ -225,15 +225,6 @@ spec:
                       type: object
                     type: array
                 type: object
-              computeIsolation:
-                default: soft
-                description: How to isolate computing resources, could be `shared`
-                  or `soft` or `hard`
-                enum:
-                - shared
-                - soft
-                - hard
-                type: string
               gpuCount:
                 description: The number of GPUs to be used by the workload, default
                   to 1
@@ -253,6 +244,16 @@ spec:
                 description: Schedule the workload to the same GPU server that runs
                   vGPU worker for best performance, default to false
                 type: boolean
+              isolation:
+                default: soft
+                description: How to isolate resources, could be `shared` or `soft`
+                  or `hard` or `partitioned`
+                enum:
+                - shared
+                - soft
+                - hard
+                - partitioned
+                type: string
               nodeAffinity:
                 description: NodeAffinity specifies the node affinity requirements
                   for the workload