X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/7be9cb0ae8aeb5a49d5450aa38ff9f652761c2d3..7000c1ebd170001e10807b583a29e9e7e9570b23:/lib/crunchrun/docker.go

diff --git a/lib/crunchrun/docker.go b/lib/crunchrun/docker.go
index c64ff83da3..e62f2a39ba 100644
--- a/lib/crunchrun/docker.go
+++ b/lib/crunchrun/docker.go
@@ -71,7 +71,7 @@ func (e *dockerExecutor) LoadImage(imageID string, imageTarballPath string, cont
 	return nil
 }
 
-func (e *dockerExecutor) Create(spec containerSpec) error {
+func (e *dockerExecutor) config(spec containerSpec) (dockercontainer.Config, dockercontainer.HostConfig) {
 	e.logf("Creating Docker container")
 	cfg := dockercontainer.Config{
 		Image:        spec.Image,
@@ -107,10 +107,40 @@ func (e *dockerExecutor) Create(spec containerSpec) error {
 		},
 	}
 	if spec.CUDADeviceCount != 0 {
+		var deviceIds []string
+		if cudaVisibleDevices := os.Getenv("CUDA_VISIBLE_DEVICES"); cudaVisibleDevices != "" {
+			// If a resource manager such as slurm or LSF told
+			// us to select specific devices we need to propagate that.
+			deviceIds = strings.Split(cudaVisibleDevices, ",")
+		}
+
+		deviceCount := spec.CUDADeviceCount
+		if len(deviceIds) > 0 {
+			// Docker won't accept both non-empty
+			// DeviceIDs and a non-zero Count
+			//
+			// (it turns out "Count" is a dumb fallback
+			// that just allocates device 0, 1, 2, ...,
+			// Count-1)
+			deviceCount = 0
+		}
+
+		// Capabilities are confusing.  The driver has generic
+		// capabilities "gpu" and "nvidia" but then there's
+		// additional capabilities "compute" and "utility"
+		// that are passed to nvidia-container-cli.
+		//
+		// "compute" means include the CUDA libraries and
+		// "utility" means include the CUDA utility programs
+		// (like nvidia-smi).
+		//
+		// https://github.com/moby/moby/blob/7b9275c0da707b030e62c96b679a976f31f929d3/daemon/nvidia_linux.go#L37
+		// https://github.com/containerd/containerd/blob/main/contrib/nvidia/nvidia.go
 		hostCfg.Resources.DeviceRequests = append(hostCfg.Resources.DeviceRequests, dockercontainer.DeviceRequest{
 			Driver:       "nvidia",
-			Count:        spec.CUDADeviceCount,
-			Capabilities: [][]string{[]string{"gpu", "nvidia", "compute"}},
+			Count:        deviceCount,
+			DeviceIDs:    deviceIds,
+			Capabilities: [][]string{[]string{"gpu", "nvidia", "compute", "utility"}},
 		})
 	}
 	for path, mount := range spec.BindMounts {
@@ -123,7 +153,11 @@ func (e *dockerExecutor) Create(spec containerSpec) error {
 	if spec.EnableNetwork {
 		hostCfg.NetworkMode = dockercontainer.NetworkMode(spec.NetworkMode)
 	}
+	return cfg, hostCfg
+}
 
+func (e *dockerExecutor) Create(spec containerSpec) error {
+	cfg, hostCfg := e.config(spec)
 	created, err := e.dockerclient.ContainerCreate(context.TODO(), &cfg, &hostCfg, nil, e.containerUUID)
 	if err != nil {
 		return fmt.Errorf("While creating container: %v", err)
@@ -181,7 +215,6 @@ func (e *dockerExecutor) Wait(ctx context.Context) (int, error) {
 	for {
 		select {
 		case waitBody := <-waitOk:
-			e.logf("Container exited with code: %v", waitBody.StatusCode)
 			// wait for stdout/stderr to complete
 			<-e.doneIO
 			return int(waitBody.StatusCode), nil