feat(scheduler): account for multiple instances of a model per server when scheduling

apis/mlops/agent/agent.proto

@@ -27,6 +27,7 @@ message ModelEventMessage {
   Event event = 5;
   string message = 6;
   uint64 availableMemoryBytes = 7;
+  ModelRuntimeInfo runtimeInfo = 8;
 }
 
 message ModelEventResponse {
@@ -92,8 +93,29 @@ message ModelOperationMessage {
 message ModelVersion {
   scheduler.Model model = 1;
   uint32 version = 2;
+  ModelRuntimeInfo runtimeInfo = 3;
 }
 
+message ModelRuntimeInfo {
+  oneof modelRuntimeInfo {
+    MLServerModelSettings mlserver = 1; 
+    TritonModelConfig triton = 2;
+    }
+}
+
+message MLServerModelSettings {
+  uint32 parallelWorkers = 1;
+}
+
+message TritonModelConfig {
+  repeated TritonCPU cpu = 1;
+}
+
+message TritonCPU {
+  uint32 instanceCount = 1;
+  }
+
+
 // [END Messages]
 
 // [START Services]

scheduler/pkg/agent/agent_debug_test.go

@@ -35,7 +35,7 @@ func setupService(numModels int, modelPrefix string, capacity int) *agentDebug {
 }
 
 func TestAgentDebugServiceSmoke(t *testing.T) {
-	//TODO break this down in proper tests
+	// TODO break this down in proper tests
 	g := NewGomegaWithT(t)
 
 	service := setupService(10, "dummy", 10)
@@ -60,6 +60,7 @@ func TestAgentDebugServiceSmoke(t *testing.T) {
 					MemoryBytes: &mem,
 				},
 			},
+			RuntimeInfo: getModelRuntimeInfo(1),
 		},
 	)
 	g.Expect(err).To(BeNil())
@@ -87,7 +88,7 @@ func TestAgentDebugServiceSmoke(t *testing.T) {
 }
 
 func TestAgentDebugEarlyStop(t *testing.T) {
-	//TODO break this down in proper tests
+	// TODO break this down in proper tests
 	g := NewGomegaWithT(t)
 
 	service := setupService(10, "dummy", 10)

scheduler/pkg/agent/client.go

@@ -615,15 +615,23 @@ func (c *Client) LoadModel(request *agent.ModelOperationMessage, timestamp int64
 	}
 	logger.Infof("Chose path %s for model %s:%d", *chosenVersionPath, modelName, modelVersion)
 
+	modelConfig, err := c.ModelRepository.GetModelRuntimeInfo(modelWithVersion)
+	if err != nil {
+		logger.Errorf("there was a problem getting the config for model: %s", modelName)
+	}
+
 	// TODO: consider whether we need the actual protos being sent to `LoadModelVersion`?
 	modifiedModelVersionRequest := getModifiedModelVersion(
 		modelWithVersion,
 		pinnedModelVersion,
 		request.GetModelVersion(),
+		modelConfig,
 	)
+
 	loaderFn := func() error {
 		return c.stateManager.LoadModelVersion(modifiedModelVersionRequest)
 	}
+
 	if err := backoffWithMaxNumRetry(loaderFn, c.settings.maxLoadRetryCount, c.settings.maxLoadElapsedTime, logger); err != nil {
 		c.sendModelEventError(modelName, modelVersion, agent.ModelEventMessage_LOAD_FAILED, err)
 		c.cleanup(modelWithVersion)
@@ -641,7 +649,8 @@ func (c *Client) LoadModel(request *agent.ModelOperationMessage, timestamp int64
 	}
 
 	logger.Infof("Load model %s:%d success", modelName, modelVersion)
-	return c.sendAgentEvent(modelName, modelVersion, agent.ModelEventMessage_LOADED)
+
+	return c.sendAgentEvent(modelName, modelVersion, modelConfig, agent.ModelEventMessage_LOADED)
 }
 
 func (c *Client) UnloadModel(request *agent.ModelOperationMessage, timestamp int64) error {
@@ -674,7 +683,7 @@ func (c *Client) UnloadModel(request *agent.ModelOperationMessage, timestamp int
 	defer c.modelTimestamps.Store(modelWithVersion, timestamp)
 
 	// we do not care about model versions here
-	modifiedModelVersionRequest := getModifiedModelVersion(modelWithVersion, pinnedModelVersion, request.GetModelVersion())
+	modifiedModelVersionRequest := getModifiedModelVersion(modelWithVersion, pinnedModelVersion, request.GetModelVersion(), nil)
 
 	unloaderFn := func() error {
 		return c.stateManager.UnloadModelVersion(modifiedModelVersionRequest)
@@ -702,7 +711,7 @@ func (c *Client) UnloadModel(request *agent.ModelOperationMessage, timestamp int
 	}
 
 	logger.Infof("Unload model %s:%d success", modelName, modelVersion)
-	return c.sendAgentEvent(modelName, modelVersion, agent.ModelEventMessage_UNLOADED)
+	return c.sendAgentEvent(modelName, modelVersion, nil, agent.ModelEventMessage_UNLOADED)
 }
 
 func (c *Client) cleanup(modelWithVersion string) {
@@ -742,6 +751,7 @@ func (c *Client) sendModelEventError(
 func (c *Client) sendAgentEvent(
 	modelName string,
 	modelVersion uint32,
+	modelRuntimeInfo *agent.ModelRuntimeInfo,
 	event agent.ModelEventMessage_Event,
 ) error {
 	// if the server is draining and the model load has succeeded, we need to "cancel"
@@ -765,6 +775,7 @@ func (c *Client) sendAgentEvent(
 		ModelVersion:         modelVersion,
 		Event:                event,
 		AvailableMemoryBytes: c.stateManager.GetAvailableMemoryBytesWithOverCommit(),
+		RuntimeInfo:          modelRuntimeInfo,
 	})
 	return err
 }

scheduler/pkg/agent/client_test.go

@@ -29,6 +29,7 @@ import (
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/client-go/kubernetes/fake"
 
+	"github.com/seldonio/seldon-core/apis/go/v2/mlops/agent"
 	pb "github.com/seldonio/seldon-core/apis/go/v2/mlops/agent"
 	pbs "github.com/seldonio/seldon-core/apis/go/v2/mlops/scheduler"
 
@@ -51,6 +52,7 @@ type mockAgentV2Server struct {
 	unloadFailedEvents int
 	otherEvents        int
 	errors             int
+	events             []*pb.ModelEventMessage
 }
 
 type FakeModelRepository struct {
@@ -64,6 +66,10 @@ func (f *FakeModelRepository) RemoveModelVersion(modelName string) error {
 	return nil
 }
 
+func (f *FakeModelRepository) GetModelRuntimeInfo(modelName string) (*pb.ModelRuntimeInfo, error) {
+	return &pb.ModelRuntimeInfo{ModelRuntimeInfo: &pb.ModelRuntimeInfo_Mlserver{Mlserver: &agent.MLServerModelSettings{ParallelWorkers: uint32(1)}}}, nil
+}
+
 func (f *FakeModelRepository) DownloadModelVersion(modelName string, version uint32, modelSpec *pbs.ModelSpec, config []byte) (*string, error) {
 	f.modelDownloads++
 	if f.err != nil {
@@ -147,6 +153,7 @@ func (m *mockAgentV2Server) AgentEvent(ctx context.Context, message *pb.ModelEve
 	default:
 		m.otherEvents++
 	}
+	m.events = append(m.events, message)
 	return &pb.ModelEventResponse{}, nil
 }
 
@@ -247,6 +254,7 @@ func TestLoadModel(t *testing.T) {
 		models                  []string
 		replicaConfig           *pb.ReplicaConfig
 		op                      *pb.ModelOperationMessage
+		modelConfig             *pb.ModelRuntimeInfo
 		expectedAvailableMemory uint64
 		v2Status                int
 		modelRepoErr            error
@@ -270,9 +278,11 @@ func TestLoadModel(t *testing.T) {
 						},
 						ModelSpec: &pbs.ModelSpec{Uri: "gs://model", MemoryBytes: &smallMemory},
 					},
+					RuntimeInfo: getModelRuntimeInfo(1),
 				},
 			},
 			replicaConfig:           &pb.ReplicaConfig{MemoryBytes: 1000},
+			modelConfig:             getModelRuntimeInfo(1),
 			expectedAvailableMemory: 500,
 			v2Status:                200,
 			success:                 true,
@@ -289,10 +299,12 @@ func TestLoadModel(t *testing.T) {
 						},
 						ModelSpec: &pbs.ModelSpec{Uri: "gs://model", MemoryBytes: &smallMemory},
 					},
+					RuntimeInfo: getModelRuntimeInfo(1),
 				},
 				AutoscalingEnabled: true,
 			},
 			replicaConfig:           &pb.ReplicaConfig{MemoryBytes: 1000},
+			modelConfig:             getModelRuntimeInfo(1),
 			expectedAvailableMemory: 500,
 			v2Status:                200,
 			success:                 true,
@@ -310,9 +322,11 @@ func TestLoadModel(t *testing.T) {
 						},
 						ModelSpec: &pbs.ModelSpec{Uri: "gs://model", MemoryBytes: &smallMemory},
 					},
+					RuntimeInfo: getModelRuntimeInfo(1),
 				},
 			},
 			replicaConfig:           &pb.ReplicaConfig{MemoryBytes: 1000},
+			modelConfig:             getModelRuntimeInfo(1),
 			expectedAvailableMemory: 1000,
 			v2Status:                400,
 			success:                 false,
@@ -329,9 +343,11 @@ func TestLoadModel(t *testing.T) {
 						},
 						ModelSpec: &pbs.ModelSpec{Uri: "gs://model", MemoryBytes: &largeMemory},
 					},
+					RuntimeInfo: getModelRuntimeInfo(1),
 				},
 			},
 			replicaConfig:           &pb.ReplicaConfig{MemoryBytes: 1000},
+			modelConfig:             getModelRuntimeInfo(1),
 			expectedAvailableMemory: 1000,
 			v2Status:                200,
 			success:                 false,
@@ -399,6 +415,9 @@ func TestLoadModel(t *testing.T) {
 				g.Expect(err).To(BeNil())
 				g.Expect(mockAgentV2Server.loadedEvents).To(Equal(1))
 				g.Expect(mockAgentV2Server.loadFailedEvents).To(Equal(0))
+				g.Expect(len(mockAgentV2Server.events)).To(Equal(1))
+				g.Expect(mockAgentV2Server.events[0].RuntimeInfo).ToNot(BeNil())
+				g.Expect(mockAgentV2Server.events[0].RuntimeInfo.GetMlserver().ParallelWorkers).To(Equal(uint32(1)))
 				g.Expect(client.stateManager.GetAvailableMemoryBytes()).To(Equal(test.expectedAvailableMemory))
 				g.Expect(modelRepository.modelRemovals).To(Equal(0))
 				loadedVersions := client.stateManager.modelVersions.getVersionsForAllModels()

scheduler/pkg/agent/client_utils.go

@@ -54,10 +54,11 @@ func isReady(service interfaces.DependencyServiceInterface, logger *log.Entry, m
 	return backoff.RetryNotify(readyToError, backoffWithMax, logFailure)
 }
 
-func getModifiedModelVersion(modelId string, version uint32, originalModelVersion *agent.ModelVersion) *agent.ModelVersion {
+func getModifiedModelVersion(modelId string, version uint32, originalModelVersion *agent.ModelVersion, modelRuntimeInfo *agent.ModelRuntimeInfo) *agent.ModelVersion {
 	mv := proto.Clone(originalModelVersion)
 	mv.(*agent.ModelVersion).Model.Meta.Name = modelId
 	mv.(*agent.ModelVersion).Version = version
+	mv.(*agent.ModelVersion).RuntimeInfo = modelRuntimeInfo
 	return mv.(*agent.ModelVersion)
 }
 

scheduler/pkg/agent/model_state.go

@@ -59,7 +59,6 @@ func (modelState *ModelState) addModelVersionImpl(modelVersionDetails *agent.Mod
 				modelName, versionId, exsistingVersion.getVersion())
 		}
 	}
-
 }
 
 // Remove model version and return true if no versions left (in which case we remove from map)
@@ -70,7 +69,6 @@ func (modelState *ModelState) removeModelVersion(modelVersionDetails *agent.Mode
 }
 
 func (modelState *ModelState) removeModelVersionImpl(modelVersionDetails *agent.ModelVersion) (bool, error) {
-
 	modelName := modelVersionDetails.GetModel().GetMeta().GetName()
 	versionId := modelVersionDetails.GetVersion()
 
@@ -143,7 +141,8 @@ func (modelState *ModelState) getVersionsForAllModels() []*agent.ModelVersion {
 		mv := version.get()
 		versionedModelName := mv.Model.GetMeta().Name
 		originalModelName, originalModelVersion, _ := util.GetOrignalModelNameAndVersion(versionedModelName)
-		loadedModels = append(loadedModels, getModifiedModelVersion(originalModelName, originalModelVersion, mv))
+		modelRuntimeInfo := mv.RuntimeInfo
+		loadedModels = append(loadedModels, getModifiedModelVersion(originalModelName, originalModelVersion, mv, modelRuntimeInfo))
 	}
 	return loadedModels
 }
@@ -153,7 +152,19 @@ type modelVersion struct {
 }
 
 func (version *modelVersion) getVersionMemory() uint64 {
-	return version.versionInfo.GetModel().GetModelSpec().GetMemoryBytes()
+	instanceCount := getInstanceCount(version)
+	return version.versionInfo.GetModel().GetModelSpec().GetMemoryBytes() * instanceCount
+}
+
+func getInstanceCount(version *modelVersion) uint64 {
+	switch version.versionInfo.RuntimeInfo.ModelRuntimeInfo.(type) {
+	case *agent.ModelRuntimeInfo_Mlserver:
+		return uint64(version.versionInfo.GetRuntimeInfo().GetMlserver().ParallelWorkers)
+	case *agent.ModelRuntimeInfo_Triton:
+		return uint64(version.versionInfo.GetRuntimeInfo().GetTriton().Cpu[0].InstanceCount)
+	default:
+		return 1
+	}
 }
 
 func (version *modelVersion) getVersion() uint32 {