docs: how to enable nvidia gpu 2.0 in Pouch

Signed-off-by: codejuan <xh@decbug.com>

daemon/mgr/spec_nvidia_hook.go

@@ -2,13 +2,15 @@ package mgr
 
 import (
 	"os/exec"
+	"path"
 
 	"github.com/alibaba/pouch/pkg/utils"
 
 	"github.com/opencontainers/runtime-spec/specs-go"
 )
 
 var (
+	// nvidiaHookName is a custom OCI prestart hook binary to runc in order to enable GPU containers.
 	nvidiaHookName = "nvidia-container-runtime-hook"
 )
 
@@ -25,13 +27,19 @@ func setNvidiaHook(c *Container, spec *SpecWrapper) error {
 		return nil
 	}
 
-	path, err := exec.LookPath(nvidiaHookName)
-	if err != nil {
-		return err
+	hookPath := ""
+	if !path.IsAbs(nvidiaHookName) {
+		var err error
+		hookPath, err = exec.LookPath(nvidiaHookName)
+		if err != nil {
+			return err
+		}
+	} else {
+		hookPath = nvidiaHookName
 	}
-	args := []string{path}
+	args := []string{hookPath}
 	nvidiaPrestart := specs.Hook{
-		Path: path,
+		Path: hookPath,
 		Args: append(args, "prestart"),
 	}
 	spec.s.Hooks.Prestart = append(spec.s.Hooks.Prestart, nvidiaPrestart)

daemon/mgr/spec_nvidia_hook_test.go

@@ -13,12 +13,18 @@ import (
 
 func Test_setNvidiaHook(t *testing.T) {
 	nvidiaHookName = "test-nvidia-container-runtime-hook"
-	installDir := "/usr/local/bin/"
+	installDir := "/usr/local/bin"
 	fullname := path.Join(installDir, nvidiaHookName)
+	nvidiaHookName = fullname
 	os.Remove(fullname)
 	os.Create(fullname)
 	os.Chmod(fullname, 0755)
-	path, _ := exec.LookPath(nvidiaHookName)
+	hookPath := ""
+	if !path.IsAbs(nvidiaHookName) {
+		hookPath, _ = exec.LookPath(nvidiaHookName)
+	} else {
+		hookPath = nvidiaHookName
+	}
 	defer func() {
 		os.Remove(fullname)
 	}()
@@ -69,10 +75,36 @@ func Test_setNvidiaHook(t *testing.T) {
 					},
 				},
 			},
-			// exec.LookPath("nvidia-container-runtime-hook") return error,
-			[]specs.Hook{specs.Hook{
-				Path: path,
-				Args: append([]string{path}, "prestart"),
+			[]specs.Hook{{
+				Path: hookPath,
+				Args: append([]string{hookPath}, "prestart"),
+			}},
+		},
+		{
+			"NvidiaConfig not nil, NvidiaEnv is null",
+			&Container{
+				HostConfig: &types.HostConfig{
+					Resources: types.Resources{
+						NvidiaConfig: &types.NvidiaConfig{
+							NvidiaDriverCapabilities: "all",
+							NvidiaVisibleDevices:     "all",
+						},
+					},
+				},
+				Config: &types.ContainerConfig{
+					Env: []string{},
+				},
+			},
+			&SpecWrapper{
+				s: &specs.Spec{
+					Hooks: &specs.Hooks{
+						Prestart: []specs.Hook{},
+					},
+				},
+			},
+			[]specs.Hook{{
+				Path: hookPath,
+				Args: append([]string{hookPath}, "prestart"),
 			}},
 		},
 	}

docs/features/pouch_with_gpu.md

@@ -0,0 +1,38 @@
+# PouchContainer with GPU
+
+NVIDIA uses containers to develop, test, benchmark, and deploy deep learning (DL) frameworks and HPC applications. A variety of customers used NVIDIA-Container-Runtime to containerize and run GPU accelerated workloads. The NVIDIA-Container-Runtime is the next-generation GPU-aware container runtime. It is compatible with the Open Containers Initiative (OCI) specification used by Pouch and other popular container technologies.
+
+## nvidia gpu runtime
+
+![pouch_with_pouch](../static_files/pouch_with_gpu.png)
+
+The figure shows how the NVIDIA-Container-Runtime integrates into Pouch. We set a custom OCI prestart hook called nvidia-container-runtime-hook to runc in order to enable GPU containers.
+
+## Prerequisites Installation
+
+Make sure you have installed the NVIDIA driver
+
+1. GNU/Linux x86_64 with kernel version > 3.10
+2. NVIDIA GPU with Architecture > Fermi (2.1)
+3. NVIDIA drivers ~= 361.93 (untested on older versions)
+
+## Start GPU container
+
+Pouch support 2 method to start GPU container
+
+1. Via nvidia config API, [nvidiaconfig](https://github.com/alibaba/pouch/blob/master/docs/api/HTTP_API.md#nvidiaconfig)
+2. Via Environment variables, [nvidia-container-runtime-env](https://github.com/NVIDIA/nvidia-container-runtime#environment-variables-oci-spec)
+
+### Via API
+
+```shell
+pouch run -it --nvidia-capabilities=all --nvidia-visible-devs=all centos:7 bash
+```
+
+### Via Environment variables
+
+The NVIDIA Container Runtime uses environment variables in container images to specify a GPU accelerated container.
+
+```shell
+pouch run -it -e NVIDIA_VISIBLE_DEVICES=all -e NVIDIA_DRIVER_CAPABILITIES=all centos:7 bash
+```