feat: add device controller

Author: Code2Life

cmd/hypervisor/main.go

@@ -36,28 +36,11 @@ func main() {
 		klog.Fatalf("Failed to start device manager: %v", err)
 	}
 	defer mgr.Stop()
-
 	klog.Info("Device manager started")
 
-	// Discover devices
 	devices := mgr.GetDevices()
-	klog.Infof("Discovered %d devices", len(devices))
-
 	if len(devices) == 0 {
-		klog.Warning("No devices discovered, waiting...")
-		time.Sleep(2 * time.Second)
-		devices = mgr.GetDevices()
-		if len(devices) == 0 {
-			klog.Fatalf("No devices available")
-		}
-	}
-
-	// Register default pool
-	deviceUUIDs := make([]string, 0, len(devices))
-	for _, d := range devices {
-		deviceUUIDs = append(deviceUUIDs, d.UUID)
-		klog.Infof("Device: UUID=%s, Vendor=%s, Model=%s, Memory=%d GB",
-			d.UUID, d.Vendor, d.Model, d.TotalMemory/(1024*1024*1024))
+		klog.Fatalf("No devices found")
 	}
 
 	// Parse isolation mode
@@ -75,31 +58,13 @@ func main() {
 		klog.Fatalf("Invalid isolation mode: %s", *isolationMode)
 	}
 
-	pool := &device.DevicePool{
-		Vendor:         devices[0].Vendor,
-		IsolationMode:  mode,
-		DeviceUUIDs:    deviceUUIDs,
-		AcceleratorLib: *acceleratorLibPath,
-	}
-
-	if err := mgr.RegisterPool(pool); err != nil {
-		klog.Fatalf("Failed to register pool: %v", err)
-	}
-	klog.Infof("Registered devices: %s with %d devices, isolation mode: %s", devices[0].Vendor, len(deviceUUIDs), mode)
-
-	// TODO: 2. If k8s mode, listen Pods from kubelet socket and build a map
-	// TODO: 3. Extensible Device Plugin, to read config yaml of pool and
-	// TODO: 4. Report GPU CR to API server, if DRA enabled, report ResourceSlice
-	// TODO: 5. Build shm handle or ivshmem device for soft isolation mode for
-	//        limiter and hard isolation mode, manage shm lifecycle
-	// TODO: 6. Expose HTTP APIs for watch worker pod status, or create workers process,
-	//  manage workers lifecycle in VM mode
+	klog.Infof("Registered devices: %s with %d devices, isolation mode: %s", devices[0].Vendor, len(devices), mode)
 
 	// Wait for interrupt signal
 	sigCh := make(chan os.Signal, 1)
 	signal.Notify(sigCh, os.Interrupt, syscall.SIGTERM)
 
-	klog.Info("Hypervisor running, press Ctrl+C to stop")
+	klog.Info("Hypervisor running")
 	<-sigCh
 	klog.Info("Shutting down...")
 }

internal/hypervisor/backend/backend.go

@@ -0,0 +1,41 @@
+package integration
+
+import (
+	"context"
+
+	"github.com/NexusGPU/tensor-fusion/internal/hypervisor/device"
+)
+
+type Framework interface {
+	AllocateDevice(ctx context.Context, request *device.DeviceAllocateRequest) (*device.DeviceAllocateResponse, error)
+
+	ListDevices(ctx context.Context) ([]*device.DeviceInfo, error)
+
+	DevicesUpdates(ctx context.Context) (<-chan []*device.DeviceInfo, error)
+
+	GetDevice(ctx context.Context, deviceUUID string) (*device.DeviceInfo, error)
+
+	GetDeviceAllocations(ctx context.Context, deviceUUID string) ([]*device.DeviceAllocation, error)
+
+	GetDeviceAllocationUpdates(ctx context.Context, deviceUUID string, allocationID string) (<-chan []*device.DeviceAllocation, error)
+}
+
+// The backend interface for the hypervisor to interact with the underlying infrastructure
+type Backend interface {
+	Start(ctx context.Context, framework Framework, params map[string]string) error
+
+	// Get GPU workers from the workload orchestration platform
+	ListAndWatchWorkers(ctx context.Context) ([]string, error)
+
+	// Report devices to backend orchestration and O&M platform
+	ReportDevices(ctx context.Context, devices []string) error
+
+	// Link workers to actual running process list on OS
+	GetWorkerProcessMap(ctx context.Context) (map[string][]string, error)
+
+	// Spawn worker process on OS
+	StartWorker(ctx context.Context, workerUID string) error
+
+	// Stop worker process on OS
+	StopWorker(ctx context.Context, workerUID string) error
+}

internal/hypervisor/backend/kubernetes/apiserver.go

@@ -0,0 +1 @@
+package kubernetes

internal/hypervisor/backend/kubernetes/deviceplugin.go

@@ -0,0 +1 @@
+package kubernetes

internal/hypervisor/backend/kubernetes/dra.go

@@ -0,0 +1 @@
+package kubernetes

internal/hypervisor/backend/kubernetes/kubelet.go

@@ -0,0 +1 @@
+package kubernetes

internal/hypervisor/backend/kubernetes/kubernetes_backend.go

@@ -0,0 +1 @@
+package kubernetes

internal/hypervisor/backend/kubernetes/ns_mapper.go

@@ -0,0 +1 @@
+package kubernetes

internal/hypervisor/backend/singlenode/filestate.go

@@ -0,0 +1 @@
+package singlenode

internal/hypervisor/backend/singlenode/singlenode_backend.go

@@ -0,0 +1 @@
+package singlenode