X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/c217491f9fdc78bd1c665618137c053a852599ac..f389a552e9b81de0e2030fe2fddc4852b33f63cc:/services/crunch-run/crunchrun.go

diff --git a/services/crunch-run/crunchrun.go b/services/crunch-run/crunchrun.go
index acaa5a2165..812525db69 100644
--- a/services/crunch-run/crunchrun.go
+++ b/services/crunch-run/crunchrun.go
@@ -1,31 +1,46 @@
 package main
 
 import (
+	"bytes"
+	"context"
 	"encoding/json"
 	"errors"
 	"flag"
 	"fmt"
-	"git.curoverse.com/arvados.git/sdk/go/arvadosclient"
-	"git.curoverse.com/arvados.git/sdk/go/keepclient"
-	"git.curoverse.com/arvados.git/sdk/go/manifest"
-	"github.com/curoverse/dockerclient"
 	"io"
 	"io/ioutil"
 	"log"
 	"os"
 	"os/exec"
 	"os/signal"
+	"path"
+	"path/filepath"
+	"sort"
 	"strings"
 	"sync"
 	"syscall"
 	"time"
+
+	"git.curoverse.com/arvados.git/lib/crunchstat"
+	"git.curoverse.com/arvados.git/sdk/go/arvados"
+	"git.curoverse.com/arvados.git/sdk/go/arvadosclient"
+	"git.curoverse.com/arvados.git/sdk/go/keepclient"
+	"git.curoverse.com/arvados.git/sdk/go/manifest"
+
+	dockertypes "github.com/docker/docker/api/types"
+	dockercontainer "github.com/docker/docker/api/types/container"
+	dockernetwork "github.com/docker/docker/api/types/network"
+	dockerclient "github.com/docker/docker/client"
 )
 
 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
 type IArvadosClient interface {
 	Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
 	Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
-	Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) (err error)
+	Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
+	Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
+	CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
+	Discovery(key string) (interface{}, error)
 }
 
 // ErrCancelled is the error returned when the container is cancelled.
@@ -34,57 +49,74 @@ var ErrCancelled = errors.New("Cancelled")
 // IKeepClient is the minimal Keep API methods used by crunch-run.
 type IKeepClient interface {
 	PutHB(hash string, buf []byte) (string, int, error)
-	ManifestFileReader(m manifest.Manifest, filename string) (keepclient.ReadCloserWithLen, error)
+	ManifestFileReader(m manifest.Manifest, filename string) (keepclient.Reader, error)
 }
 
-// Mount describes the mount points to create inside the container.
-type Mount struct {
-	Kind             string `json:"kind"`
-	Writable         bool   `json:"writable"`
-	PortableDataHash string `json:"portable_data_hash"`
-	UUID             string `json:"uuid"`
-	DeviceType       string `json:"device_type"`
-	Path             string `json:"path"`
+// NewLogWriter is a factory function to create a new log writer.
+type NewLogWriter func(name string) io.WriteCloser
+
+type RunArvMount func(args []string, tok string) (*exec.Cmd, error)
+
+type MkTempDir func(string, string) (string, error)
+
+// ThinDockerClient is the minimal Docker client interface used by crunch-run.
+type ThinDockerClient interface {
+	ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error)
+	ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
+		networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error)
+	ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error
+	ContainerStop(ctx context.Context, container string, timeout *time.Duration) error
+	ContainerWait(ctx context.Context, container string) (int64, error)
+	ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error)
+	ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error)
+	ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error)
 }
 
-// Collection record returned by the API server.
-type CollectionRecord struct {
-	ManifestText     string `json:"manifest_text"`
-	PortableDataHash string `json:"portable_data_hash"`
+// ThinDockerClientProxy is a proxy implementation of ThinDockerClient
+// that executes the docker requests on dockerclient.Client
+type ThinDockerClientProxy struct {
+	Docker *dockerclient.Client
 }
 
-// ContainerRecord is the container record returned by the API server.
-type ContainerRecord struct {
-	UUID               string                 `json:"uuid"`
-	Command            []string               `json:"command"`
-	ContainerImage     string                 `json:"container_image"`
-	Cwd                string                 `json:"cwd"`
-	Environment        map[string]string      `json:"environment"`
-	Mounts             map[string]Mount       `json:"mounts"`
-	OutputPath         string                 `json:"output_path"`
-	Priority           int                    `json:"priority"`
-	RuntimeConstraints map[string]interface{} `json:"runtime_constraints"`
-	State              string                 `json:"state"`
-	Output             string                 `json:"output"`
+// ContainerAttach invokes dockerclient.Client.ContainerAttach
+func (proxy ThinDockerClientProxy) ContainerAttach(ctx context.Context, container string, options dockertypes.ContainerAttachOptions) (dockertypes.HijackedResponse, error) {
+	return proxy.Docker.ContainerAttach(ctx, container, options)
 }
 
-// NewLogWriter is a factory function to create a new log writer.
-type NewLogWriter func(name string) io.WriteCloser
+// ContainerCreate invokes dockerclient.Client.ContainerCreate
+func (proxy ThinDockerClientProxy) ContainerCreate(ctx context.Context, config *dockercontainer.Config, hostConfig *dockercontainer.HostConfig,
+	networkingConfig *dockernetwork.NetworkingConfig, containerName string) (dockercontainer.ContainerCreateCreatedBody, error) {
+	return proxy.Docker.ContainerCreate(ctx, config, hostConfig, networkingConfig, containerName)
+}
 
-type RunArvMount func([]string) (*exec.Cmd, error)
+// ContainerStart invokes dockerclient.Client.ContainerStart
+func (proxy ThinDockerClientProxy) ContainerStart(ctx context.Context, container string, options dockertypes.ContainerStartOptions) error {
+	return proxy.Docker.ContainerStart(ctx, container, options)
+}
 
-type MkTempDir func(string, string) (string, error)
+// ContainerStop invokes dockerclient.Client.ContainerStop
+func (proxy ThinDockerClientProxy) ContainerStop(ctx context.Context, container string, timeout *time.Duration) error {
+	return proxy.Docker.ContainerStop(ctx, container, timeout)
+}
 
-// ThinDockerClient is the minimal Docker client interface used by crunch-run.
-type ThinDockerClient interface {
-	StopContainer(id string, timeout int) error
-	InspectImage(id string) (*dockerclient.ImageInfo, error)
-	LoadImage(reader io.Reader) error
-	CreateContainer(config *dockerclient.ContainerConfig, name string, authConfig *dockerclient.AuthConfig) (string, error)
-	StartContainer(id string, config *dockerclient.HostConfig) error
-	AttachContainer(id string, options *dockerclient.AttachOptions) (io.ReadCloser, error)
-	Wait(id string) <-chan dockerclient.WaitResult
-	RemoveImage(name string, force bool) ([]*dockerclient.ImageDelete, error)
+// ContainerWait invokes dockerclient.Client.ContainerWait
+func (proxy ThinDockerClientProxy) ContainerWait(ctx context.Context, container string) (int64, error) {
+	return proxy.Docker.ContainerWait(ctx, container)
+}
+
+// ImageInspectWithRaw invokes dockerclient.Client.ImageInspectWithRaw
+func (proxy ThinDockerClientProxy) ImageInspectWithRaw(ctx context.Context, image string) (dockertypes.ImageInspect, []byte, error) {
+	return proxy.Docker.ImageInspectWithRaw(ctx, image)
+}
+
+// ImageLoad invokes dockerclient.Client.ImageLoad
+func (proxy ThinDockerClientProxy) ImageLoad(ctx context.Context, input io.Reader, quiet bool) (dockertypes.ImageLoadResponse, error) {
+	return proxy.Docker.ImageLoad(ctx, input, quiet)
+}
+
+// ImageRemove invokes dockerclient.Client.ImageRemove
+func (proxy ThinDockerClientProxy) ImageRemove(ctx context.Context, image string, options dockertypes.ImageRemoveOptions) ([]dockertypes.ImageDeleteResponseItem, error) {
+	return proxy.Docker.ImageRemove(ctx, image, options)
 }
 
 // ContainerRunner is the main stateful struct used for a single execution of a
@@ -93,15 +125,17 @@ type ContainerRunner struct {
 	Docker    ThinDockerClient
 	ArvClient IArvadosClient
 	Kc        IKeepClient
-	ContainerRecord
-	dockerclient.ContainerConfig
+	arvados.Container
+	ContainerConfig dockercontainer.Config
+	dockercontainer.HostConfig
+	token       string
 	ContainerID string
 	ExitCode    *int
 	NewLogWriter
 	loggingDone   chan bool
 	CrunchLog     *ThrottledLogger
 	Stdout        io.WriteCloser
-	Stderr        *ThrottledLogger
+	Stderr        io.WriteCloser
 	LogCollection *CollectionWriter
 	LogsPDH       *string
 	RunArvMount
@@ -111,12 +145,36 @@ type ContainerRunner struct {
 	HostOutputDir  string
 	CleanupTempDir []string
 	Binds          []string
+	Volumes        map[string]struct{}
 	OutputPDH      *string
-	CancelLock     sync.Mutex
-	Cancelled      bool
 	SigChan        chan os.Signal
 	ArvMountExit   chan error
 	finalState     string
+
+	statLogger   io.WriteCloser
+	statReporter *crunchstat.Reporter
+	statInterval time.Duration
+	cgroupRoot   string
+	// What we expect the container's cgroup parent to be.
+	expectCgroupParent string
+	// What we tell docker to use as the container's cgroup
+	// parent. Note: Ideally we would use the same field for both
+	// expectCgroupParent and setCgroupParent, and just make it
+	// default to "docker". However, when using docker < 1.10 with
+	// systemd, specifying a non-empty cgroup parent (even the
+	// default value "docker") hits a docker bug
+	// (https://github.com/docker/docker/issues/17126). Using two
+	// separate fields makes it possible to use the "expect cgroup
+	// parent to be X" feature even on sites where the "specify
+	// cgroup parent" feature breaks.
+	setCgroupParent string
+
+	cStateLock sync.Mutex
+	cStarted   bool // StartContainer() succeeded
+	cCancelled bool // StopContainer() invoked
+
+	enableNetwork string // one of "default" or "always"
+	networkMode   string // passed through to HostConfig.NetworkMode
 }
 
 // SetupSignals sets up signal handling to gracefully terminate the underlying
@@ -127,29 +185,39 @@ func (runner *ContainerRunner) SetupSignals() {
 	signal.Notify(runner.SigChan, syscall.SIGINT)
 	signal.Notify(runner.SigChan, syscall.SIGQUIT)
 
-	go func(sig <-chan os.Signal) {
-		for _ = range sig {
-			if !runner.Cancelled {
-				runner.CancelLock.Lock()
-				runner.Cancelled = true
-				if runner.ContainerID != "" {
-					runner.Docker.StopContainer(runner.ContainerID, 10)
-				}
-				runner.CancelLock.Unlock()
-			}
-		}
+	go func(sig chan os.Signal) {
+		<-sig
+		runner.stop()
+		signal.Stop(sig)
 	}(runner.SigChan)
 }
 
+// stop the underlying Docker container.
+func (runner *ContainerRunner) stop() {
+	runner.cStateLock.Lock()
+	defer runner.cStateLock.Unlock()
+	if runner.cCancelled {
+		return
+	}
+	runner.cCancelled = true
+	if runner.cStarted {
+		timeout := time.Duration(10)
+		err := runner.Docker.ContainerStop(context.TODO(), runner.ContainerID, &(timeout))
+		if err != nil {
+			log.Printf("StopContainer failed: %s", err)
+		}
+	}
+}
+
 // LoadImage determines the docker image id from the container record and
 // checks if it is available in the local Docker image store.  If not, it loads
 // the image from Keep.
 func (runner *ContainerRunner) LoadImage() (err error) {
 
-	runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.ContainerRecord.ContainerImage)
+	runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 
-	var collection CollectionRecord
-	err = runner.ArvClient.Get("collections", runner.ContainerRecord.ContainerImage, nil, &collection)
+	var collection arvados.Collection
+	err = runner.ArvClient.Get("collections", runner.Container.ContainerImage, nil, &collection)
 	if err != nil {
 		return fmt.Errorf("While getting container image collection: %v", err)
 	}
@@ -165,7 +233,7 @@ func (runner *ContainerRunner) LoadImage() (err error) {
 
 	runner.CrunchLog.Printf("Using Docker image id '%s'", imageID)
 
-	_, err = runner.Docker.InspectImage(imageID)
+	_, _, err = runner.Docker.ImageInspectWithRaw(context.TODO(), imageID)
 	if err != nil {
 		runner.CrunchLog.Print("Loading Docker image from keep")
 
@@ -175,10 +243,11 @@ func (runner *ContainerRunner) LoadImage() (err error) {
 			return fmt.Errorf("While creating ManifestFileReader for container image: %v", err)
 		}
 
-		err = runner.Docker.LoadImage(readCloser)
+		response, err := runner.Docker.ImageLoad(context.TODO(), readCloser, false)
 		if err != nil {
 			return fmt.Errorf("While loading container image into Docker: %v", err)
 		}
+		response.Body.Close()
 	} else {
 		runner.CrunchLog.Print("Docker image is available")
 	}
@@ -188,8 +257,19 @@ func (runner *ContainerRunner) LoadImage() (err error) {
 	return nil
 }
 
-func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string) (c *exec.Cmd, err error) {
+func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string, token string) (c *exec.Cmd, err error) {
 	c = exec.Command("arv-mount", arvMountCmd...)
+
+	// Copy our environment, but override ARVADOS_API_TOKEN with
+	// the container auth token.
+	c.Env = nil
+	for _, s := range os.Environ() {
+		if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
+			c.Env = append(c.Env, s)
+		}
+	}
+	c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
+
 	nt := NewThrottledLogger(runner.NewLogWriter("arv-mount"))
 	c.Stdout = nt
 	c.Stderr = nt
@@ -232,8 +312,15 @@ func (runner *ContainerRunner) ArvMountCmd(arvMountCmd []string) (c *exec.Cmd, e
 	return c, nil
 }
 
+func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
+	if runner.ArvMountPoint == "" {
+		runner.ArvMountPoint, err = runner.MkTempDir("", prefix)
+	}
+	return
+}
+
 func (runner *ContainerRunner) SetupMounts() (err error) {
-	runner.ArvMountPoint, err = runner.MkTempDir("", "keep")
+	err = runner.SetupArvMountPoint("keep")
 	if err != nil {
 		return fmt.Errorf("While creating keep mount temp dir: %v", err)
 	}
@@ -243,27 +330,59 @@ func (runner *ContainerRunner) SetupMounts() (err error) {
 	pdhOnly := true
 	tmpcount := 0
 	arvMountCmd := []string{"--foreground", "--allow-other", "--read-write"}
+
+	if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
+		arvMountCmd = append(arvMountCmd, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
+	}
+
 	collectionPaths := []string{}
 	runner.Binds = nil
+	runner.Volumes = make(map[string]struct{})
+	needCertMount := true
+
+	var binds []string
+	for bind, _ := range runner.Container.Mounts {
+		binds = append(binds, bind)
+	}
+	sort.Strings(binds)
 
-	for bind, mnt := range runner.ContainerRecord.Mounts {
-		if bind == "stdout" {
+	for _, bind := range binds {
+		mnt := runner.Container.Mounts[bind]
+		if bind == "stdout" || bind == "stderr" {
 			// Is it a "file" mount kind?
 			if mnt.Kind != "file" {
-				return fmt.Errorf("Unsupported mount kind '%s' for stdout. Only 'file' is supported.", mnt.Kind)
+				return fmt.Errorf("Unsupported mount kind '%s' for %s. Only 'file' is supported.", mnt.Kind, bind)
 			}
 
 			// Does path start with OutputPath?
-			prefix := runner.ContainerRecord.OutputPath
+			prefix := runner.Container.OutputPath
 			if !strings.HasSuffix(prefix, "/") {
 				prefix += "/"
 			}
 			if !strings.HasPrefix(mnt.Path, prefix) {
-				return fmt.Errorf("Stdout path does not start with OutputPath: %s, %s", mnt.Path, prefix)
+				return fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
+			}
+		}
+
+		if bind == "stdin" {
+			// Is it a "collection" mount kind?
+			if mnt.Kind != "collection" && mnt.Kind != "json" {
+				return fmt.Errorf("Unsupported mount kind '%s' for stdin. Only 'collection' or 'json' are supported.", mnt.Kind)
+			}
+		}
+
+		if bind == "/etc/arvados/ca-certificates.crt" {
+			needCertMount = false
+		}
+
+		if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
+			if mnt.Kind != "collection" {
+				return fmt.Errorf("Only mount points of kind 'collection' are supported underneath the output_path: %v", bind)
 			}
 		}
 
-		if mnt.Kind == "collection" {
+		switch {
+		case mnt.Kind == "collection" && bind != "stdin":
 			var src string
 			if mnt.UUID != "" && mnt.PortableDataHash != "" {
 				return fmt.Errorf("Cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
@@ -278,7 +397,21 @@ func (runner *ContainerRunner) SetupMounts() (err error) {
 				if mnt.Writable {
 					return fmt.Errorf("Can never write to a collection specified by portable data hash")
 				}
+				idx := strings.Index(mnt.PortableDataHash, "/")
+				if idx > 0 {
+					mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
+					mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
+					runner.Container.Mounts[bind] = mnt
+				}
 				src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
+				if mnt.Path != "" && mnt.Path != "." {
+					if strings.HasPrefix(mnt.Path, "./") {
+						mnt.Path = mnt.Path[2:]
+					} else if strings.HasPrefix(mnt.Path, "/") {
+						mnt.Path = mnt.Path[1:]
+					}
+					src += "/" + mnt.Path
+				}
 			} else {
 				src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 				arvMountCmd = append(arvMountCmd, "--mount-tmp")
@@ -286,44 +419,57 @@ func (runner *ContainerRunner) SetupMounts() (err error) {
 				tmpcount += 1
 			}
 			if mnt.Writable {
-				if bind == runner.ContainerRecord.OutputPath {
+				if bind == runner.Container.OutputPath {
 					runner.HostOutputDir = src
+				} else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
+					return fmt.Errorf("Writable mount points are not permitted underneath the output_path: %v", bind)
 				}
 				runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", src, bind))
 			} else {
 				runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", src, bind))
 			}
 			collectionPaths = append(collectionPaths, src)
-		} else if mnt.Kind == "tmp" {
-			if bind == runner.ContainerRecord.OutputPath {
-				runner.HostOutputDir, err = runner.MkTempDir("", "")
-				if err != nil {
-					return fmt.Errorf("While creating mount temp dir: %v", err)
-				}
-				st, staterr := os.Stat(runner.HostOutputDir)
-				if staterr != nil {
-					return fmt.Errorf("While Stat on temp dir: %v", staterr)
-				}
-				err = os.Chmod(runner.HostOutputDir, st.Mode()|os.ModeSetgid|0777)
-				if staterr != nil {
-					return fmt.Errorf("While Chmod temp dir: %v", err)
-				}
-				runner.CleanupTempDir = append(runner.CleanupTempDir, runner.HostOutputDir)
-				runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", runner.HostOutputDir, bind))
-			} else {
-				runner.Binds = append(runner.Binds, bind)
+
+		case mnt.Kind == "tmp" && bind == runner.Container.OutputPath:
+			runner.HostOutputDir, err = runner.MkTempDir("", "")
+			if err != nil {
+				return fmt.Errorf("While creating mount temp dir: %v", err)
 			}
-		} else if mnt.Kind == "file" {
-			runner.HostOutputDir = runner.ContainerRecord.OutputPath
 			st, staterr := os.Stat(runner.HostOutputDir)
 			if staterr != nil {
-				return fmt.Errorf("While getting stat on output_path %v: %v", runner.HostOutputDir, staterr)
+				return fmt.Errorf("While Stat on temp dir: %v", staterr)
 			}
-			if st.IsDir() != true {
-				return fmt.Errorf("Given output_path '%v' is not a directory", runner.HostOutputDir)
+			err = os.Chmod(runner.HostOutputDir, st.Mode()|os.ModeSetgid|0777)
+			if staterr != nil {
+				return fmt.Errorf("While Chmod temp dir: %v", err)
 			}
-		} else {
-			return fmt.Errorf("Unknown mount kind '%s'", mnt.Kind)
+			runner.CleanupTempDir = append(runner.CleanupTempDir, runner.HostOutputDir)
+			runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s", runner.HostOutputDir, bind))
+
+		case mnt.Kind == "tmp":
+			runner.Volumes[bind] = struct{}{}
+
+		case mnt.Kind == "json":
+			jsondata, err := json.Marshal(mnt.Content)
+			if err != nil {
+				return fmt.Errorf("encoding json data: %v", err)
+			}
+			// Create a tempdir with a single file
+			// (instead of just a tempfile): this way we
+			// can ensure the file is world-readable
+			// inside the container, without having to
+			// make it world-readable on the docker host.
+			tmpdir, err := runner.MkTempDir("", "")
+			if err != nil {
+				return fmt.Errorf("creating temp dir: %v", err)
+			}
+			runner.CleanupTempDir = append(runner.CleanupTempDir, tmpdir)
+			tmpfn := filepath.Join(tmpdir, "mountdata.json")
+			err = ioutil.WriteFile(tmpfn, jsondata, 0644)
+			if err != nil {
+				return fmt.Errorf("writing temp file: %v", err)
+			}
+			runner.Binds = append(runner.Binds, fmt.Sprintf("%s:%s:ro", tmpfn, bind))
 		}
 	}
 
@@ -331,6 +477,16 @@ func (runner *ContainerRunner) SetupMounts() (err error) {
 		return fmt.Errorf("Output path does not correspond to a writable mount point")
 	}
 
+	if wantAPI := runner.Container.RuntimeConstraints.API; needCertMount && wantAPI != nil && *wantAPI {
+		for _, certfile := range arvadosclient.CertFiles {
+			_, err := os.Stat(certfile)
+			if err == nil {
+				runner.Binds = append(runner.Binds, fmt.Sprintf("%s:/etc/arvados/ca-certificates.crt:ro", certfile))
+				break
+			}
+		}
+	}
+
 	if pdhOnly {
 		arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id")
 	} else {
@@ -338,7 +494,12 @@ func (runner *ContainerRunner) SetupMounts() (err error) {
 	}
 	arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 
-	runner.ArvMount, err = runner.RunArvMount(arvMountCmd)
+	token, err := runner.ContainerToken()
+	if err != nil {
+		return fmt.Errorf("could not get container token: %s", err)
+	}
+
+	runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 	if err != nil {
 		return fmt.Errorf("While trying to start arv-mount: %v", err)
 	}
@@ -387,6 +548,14 @@ func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 				runner.CrunchLog.Printf("While closing stderr logs: %v", closeerr)
 			}
 
+			if runner.statReporter != nil {
+				runner.statReporter.Stop()
+				closeerr = runner.statLogger.Close()
+				if closeerr != nil {
+					runner.CrunchLog.Printf("While closing crunchstat logs: %v", closeerr)
+				}
+			}
+
 			runner.loggingDone <- true
 			close(runner.loggingDone)
 			return
@@ -394,98 +563,299 @@ func (runner *ContainerRunner) ProcessDockerAttach(containerReader io.Reader) {
 	}
 }
 
-// AttachLogs connects the docker container stdout and stderr logs to the
-// Arvados logger which logs to Keep and the API server logs table.
+func (runner *ContainerRunner) StartCrunchstat() {
+	runner.statLogger = NewThrottledLogger(runner.NewLogWriter("crunchstat"))
+	runner.statReporter = &crunchstat.Reporter{
+		CID:          runner.ContainerID,
+		Logger:       log.New(runner.statLogger, "", 0),
+		CgroupParent: runner.expectCgroupParent,
+		CgroupRoot:   runner.cgroupRoot,
+		PollPeriod:   runner.statInterval,
+	}
+	runner.statReporter.Start()
+}
+
+type infoCommand struct {
+	label string
+	cmd   []string
+}
+
+// Gather node information and store it on the log for debugging
+// purposes.
+func (runner *ContainerRunner) LogNodeInfo() (err error) {
+	w := runner.NewLogWriter("node-info")
+	logger := log.New(w, "node-info", 0)
+
+	commands := []infoCommand{
+		infoCommand{
+			label: "Host Information",
+			cmd:   []string{"uname", "-a"},
+		},
+		infoCommand{
+			label: "CPU Information",
+			cmd:   []string{"cat", "/proc/cpuinfo"},
+		},
+		infoCommand{
+			label: "Memory Information",
+			cmd:   []string{"cat", "/proc/meminfo"},
+		},
+		infoCommand{
+			label: "Disk Space",
+			cmd:   []string{"df", "-m", "/", os.TempDir()},
+		},
+		infoCommand{
+			label: "Disk INodes",
+			cmd:   []string{"df", "-i", "/", os.TempDir()},
+		},
+	}
+
+	// Run commands with informational output to be logged.
+	var out []byte
+	for _, command := range commands {
+		out, err = exec.Command(command.cmd[0], command.cmd[1:]...).CombinedOutput()
+		if err != nil {
+			return fmt.Errorf("While running command %q: %v",
+				command.cmd, err)
+		}
+		logger.Println(command.label)
+		for _, line := range strings.Split(string(out), "\n") {
+			logger.Println(" ", line)
+		}
+	}
+
+	err = w.Close()
+	if err != nil {
+		return fmt.Errorf("While closing node-info logs: %v", err)
+	}
+	return nil
+}
+
+// Get and save the raw JSON container record from the API server
+func (runner *ContainerRunner) LogContainerRecord() (err error) {
+	w := &ArvLogWriter{
+		ArvClient:     runner.ArvClient,
+		UUID:          runner.Container.UUID,
+		loggingStream: "container",
+		writeCloser:   runner.LogCollection.Open("container.json"),
+	}
+
+	// Get Container record JSON from the API Server
+	reader, err := runner.ArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
+	if err != nil {
+		return fmt.Errorf("While retrieving container record from the API server: %v", err)
+	}
+	defer reader.Close()
+	// Read the API server response as []byte
+	json_bytes, err := ioutil.ReadAll(reader)
+	if err != nil {
+		return fmt.Errorf("While reading container record API server response: %v", err)
+	}
+	// Decode the JSON []byte
+	var cr map[string]interface{}
+	if err = json.Unmarshal(json_bytes, &cr); err != nil {
+		return fmt.Errorf("While decoding the container record JSON response: %v", err)
+	}
+	// Re-encode it using indentation to improve readability
+	enc := json.NewEncoder(w)
+	enc.SetIndent("", "    ")
+	if err = enc.Encode(cr); err != nil {
+		return fmt.Errorf("While logging the JSON container record: %v", err)
+	}
+	err = w.Close()
+	if err != nil {
+		return fmt.Errorf("While closing container.json log: %v", err)
+	}
+	return nil
+}
+
+// AttachStreams connects the docker container stdin, stdout and stderr logs
+// to the Arvados logger which logs to Keep and the API server logs table.
 func (runner *ContainerRunner) AttachStreams() (err error) {
 
 	runner.CrunchLog.Print("Attaching container streams")
 
-	var containerReader io.Reader
-	containerReader, err = runner.Docker.AttachContainer(runner.ContainerID,
-		&dockerclient.AttachOptions{Stream: true, Stdout: true, Stderr: true})
+	// If stdin mount is provided, attach it to the docker container
+	var stdinRdr keepclient.Reader
+	var stdinJson []byte
+	if stdinMnt, ok := runner.Container.Mounts["stdin"]; ok {
+		if stdinMnt.Kind == "collection" {
+			var stdinColl arvados.Collection
+			collId := stdinMnt.UUID
+			if collId == "" {
+				collId = stdinMnt.PortableDataHash
+			}
+			err = runner.ArvClient.Get("collections", collId, nil, &stdinColl)
+			if err != nil {
+				return fmt.Errorf("While getting stding collection: %v", err)
+			}
+
+			stdinRdr, err = runner.Kc.ManifestFileReader(manifest.Manifest{Text: stdinColl.ManifestText}, stdinMnt.Path)
+			if os.IsNotExist(err) {
+				return fmt.Errorf("stdin collection path not found: %v", stdinMnt.Path)
+			} else if err != nil {
+				return fmt.Errorf("While getting stdin collection path %v: %v", stdinMnt.Path, err)
+			}
+		} else if stdinMnt.Kind == "json" {
+			stdinJson, err = json.Marshal(stdinMnt.Content)
+			if err != nil {
+				return fmt.Errorf("While encoding stdin json data: %v", err)
+			}
+		}
+	}
+
+	stdinUsed := stdinRdr != nil || len(stdinJson) != 0
+	response, err := runner.Docker.ContainerAttach(context.TODO(), runner.ContainerID,
+		dockertypes.ContainerAttachOptions{Stream: true, Stdin: stdinUsed, Stdout: true, Stderr: true})
 	if err != nil {
 		return fmt.Errorf("While attaching container stdout/stderr streams: %v", err)
 	}
 
 	runner.loggingDone = make(chan bool)
 
-	var stdoutMnt Mount
-	for bind, mnt := range runner.ContainerRecord.Mounts {
-		if bind == "stdout" {
-			stdoutMnt = mnt
-			break
-		}
-	}
-	if stdoutMnt.Path != "" {
-		stdoutPath := stdoutMnt.Path[len(runner.ContainerRecord.OutputPath):]
-		index := strings.LastIndex(stdoutPath, "/")
-		if index > 0 {
-			stdoutSubdirs := stdoutPath[:index]
-			if stdoutSubdirs != "" {
-				st, err := os.Stat(runner.HostOutputDir)
-				if err != nil {
-					return fmt.Errorf("While Stat on temp dir: %v", err)
-				}
-				path := runner.HostOutputDir + stdoutSubdirs
-				err = os.MkdirAll(path, st.Mode()|os.ModeSetgid|0777)
-				if err != nil {
-					return fmt.Errorf("While MkdirAll %q: %v", path, err)
-				}
-				st, err = os.Stat(path)
-			}
-		}
-		stdoutFile, err := os.Create(runner.HostOutputDir + "/" + stdoutPath)
+	if stdoutMnt, ok := runner.Container.Mounts["stdout"]; ok {
+		stdoutFile, err := runner.getStdoutFile(stdoutMnt.Path)
 		if err != nil {
-			return fmt.Errorf("While creating stdout file: %v", err)
+			return err
 		}
 		runner.Stdout = stdoutFile
 	} else {
 		runner.Stdout = NewThrottledLogger(runner.NewLogWriter("stdout"))
 	}
-	runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
 
-	go runner.ProcessDockerAttach(containerReader)
+	if stderrMnt, ok := runner.Container.Mounts["stderr"]; ok {
+		stderrFile, err := runner.getStdoutFile(stderrMnt.Path)
+		if err != nil {
+			return err
+		}
+		runner.Stderr = stderrFile
+	} else {
+		runner.Stderr = NewThrottledLogger(runner.NewLogWriter("stderr"))
+	}
+
+	if stdinRdr != nil {
+		go func() {
+			_, err := io.Copy(response.Conn, stdinRdr)
+			if err != nil {
+				runner.CrunchLog.Print("While writing stdin collection to docker container %q", err)
+				runner.stop()
+			}
+			stdinRdr.Close()
+			response.CloseWrite()
+		}()
+	} else if len(stdinJson) != 0 {
+		go func() {
+			_, err := io.Copy(response.Conn, bytes.NewReader(stdinJson))
+			if err != nil {
+				runner.CrunchLog.Print("While writing stdin json to docker container %q", err)
+				runner.stop()
+			}
+			response.CloseWrite()
+		}()
+	}
+
+	go runner.ProcessDockerAttach(response.Reader)
 
 	return nil
 }
 
-// StartContainer creates the container and runs it.
-func (runner *ContainerRunner) StartContainer() (err error) {
+func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
+	stdoutPath := mntPath[len(runner.Container.OutputPath):]
+	index := strings.LastIndex(stdoutPath, "/")
+	if index > 0 {
+		subdirs := stdoutPath[:index]
+		if subdirs != "" {
+			st, err := os.Stat(runner.HostOutputDir)
+			if err != nil {
+				return nil, fmt.Errorf("While Stat on temp dir: %v", err)
+			}
+			stdoutPath := path.Join(runner.HostOutputDir, subdirs)
+			err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
+			if err != nil {
+				return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
+			}
+		}
+	}
+	stdoutFile, err := os.Create(path.Join(runner.HostOutputDir, stdoutPath))
+	if err != nil {
+		return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
+	}
+
+	return stdoutFile, nil
+}
+
+// CreateContainer creates the docker container.
+func (runner *ContainerRunner) CreateContainer() error {
 	runner.CrunchLog.Print("Creating Docker container")
 
-	runner.CancelLock.Lock()
-	defer runner.CancelLock.Unlock()
+	runner.ContainerConfig.Cmd = runner.Container.Command
+	if runner.Container.Cwd != "." {
+		runner.ContainerConfig.WorkingDir = runner.Container.Cwd
+	}
 
-	if runner.Cancelled {
-		return ErrCancelled
+	for k, v := range runner.Container.Environment {
+		runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
 	}
 
-	runner.ContainerConfig.Cmd = runner.ContainerRecord.Command
-	if runner.ContainerRecord.Cwd != "." {
-		runner.ContainerConfig.WorkingDir = runner.ContainerRecord.Cwd
+	runner.ContainerConfig.Volumes = runner.Volumes
+
+	runner.HostConfig = dockercontainer.HostConfig{
+		Binds:  runner.Binds,
+		Cgroup: dockercontainer.CgroupSpec(runner.setCgroupParent),
+		LogConfig: dockercontainer.LogConfig{
+			Type: "none",
+		},
 	}
-	for k, v := range runner.ContainerRecord.Environment {
-		runner.ContainerConfig.Env = append(runner.ContainerConfig.Env, k+"="+v)
+
+	if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
+		tok, err := runner.ContainerToken()
+		if err != nil {
+			return err
+		}
+		runner.ContainerConfig.Env = append(runner.ContainerConfig.Env,
+			"ARVADOS_API_TOKEN="+tok,
+			"ARVADOS_API_HOST="+os.Getenv("ARVADOS_API_HOST"),
+			"ARVADOS_API_HOST_INSECURE="+os.Getenv("ARVADOS_API_HOST_INSECURE"),
+		)
+		runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
+	} else {
+		if runner.enableNetwork == "always" {
+			runner.HostConfig.NetworkMode = dockercontainer.NetworkMode(runner.networkMode)
+		} else {
+			runner.HostConfig.NetworkMode = dockercontainer.NetworkMode("none")
+		}
 	}
-	runner.ContainerConfig.NetworkDisabled = true
-	runner.ContainerID, err = runner.Docker.CreateContainer(&runner.ContainerConfig, "", nil)
+
+	_, stdinUsed := runner.Container.Mounts["stdin"]
+	runner.ContainerConfig.OpenStdin = stdinUsed
+	runner.ContainerConfig.StdinOnce = stdinUsed
+	runner.ContainerConfig.AttachStdin = stdinUsed
+	runner.ContainerConfig.AttachStdout = true
+	runner.ContainerConfig.AttachStderr = true
+
+	createdBody, err := runner.Docker.ContainerCreate(context.TODO(), &runner.ContainerConfig, &runner.HostConfig, nil, runner.Container.UUID)
 	if err != nil {
 		return fmt.Errorf("While creating container: %v", err)
 	}
-	hostConfig := &dockerclient.HostConfig{Binds: runner.Binds,
-		LogConfig: dockerclient.LogConfig{Type: "none"}}
 
-	err = runner.AttachStreams()
-	if err != nil {
-		return err
-	}
+	runner.ContainerID = createdBody.ID
+
+	return runner.AttachStreams()
+}
 
+// StartContainer starts the docker container created by CreateContainer.
+func (runner *ContainerRunner) StartContainer() error {
 	runner.CrunchLog.Printf("Starting Docker container id '%s'", runner.ContainerID)
-	err = runner.Docker.StartContainer(runner.ContainerID, hostConfig)
+	runner.cStateLock.Lock()
+	defer runner.cStateLock.Unlock()
+	if runner.cCancelled {
+		return ErrCancelled
+	}
+	err := runner.Docker.ContainerStart(context.TODO(), runner.ContainerID,
+		dockertypes.ContainerStartOptions{})
 	if err != nil {
-		return fmt.Errorf("While starting container: %v", err)
+		return fmt.Errorf("could not start container: %v", err)
 	}
-
+	runner.cStarted = true
 	return nil
 }
 
@@ -494,12 +864,23 @@ func (runner *ContainerRunner) StartContainer() (err error) {
 func (runner *ContainerRunner) WaitFinish() error {
 	runner.CrunchLog.Print("Waiting for container to finish")
 
-	result := runner.Docker.Wait(runner.ContainerID)
-	wr := <-result
-	if wr.Error != nil {
-		return fmt.Errorf("While waiting for container to finish: %v", wr.Error)
+	waitDocker, err := runner.Docker.ContainerWait(context.TODO(), runner.ContainerID)
+	if err != nil {
+		return fmt.Errorf("container wait: %v", err)
+	}
+
+	runner.CrunchLog.Printf("Container exited with code: %v", waitDocker)
+	code := int(waitDocker)
+	runner.ExitCode = &code
+
+	waitMount := runner.ArvMountExit
+	select {
+	case err := <-waitMount:
+		runner.CrunchLog.Printf("arv-mount exited before container finished: %v", err)
+		waitMount = nil
+		runner.stop()
+	default:
 	}
-	runner.ExitCode = &wr.ExitCode
 
 	// wait for stdout/stderr to complete
 	<-runner.loggingDone
@@ -513,6 +894,21 @@ func (runner *ContainerRunner) CaptureOutput() error {
 		return nil
 	}
 
+	if wantAPI := runner.Container.RuntimeConstraints.API; wantAPI != nil && *wantAPI {
+		// Output may have been set directly by the container, so
+		// refresh the container record to check.
+		err := runner.ArvClient.Get("containers", runner.Container.UUID,
+			nil, &runner.Container)
+		if err != nil {
+			return err
+		}
+		if runner.Container.Output != "" {
+			// Container output is already set.
+			runner.OutputPDH = &runner.Container.Output
+			return nil
+		}
+	}
+
 	if runner.HostOutputDir == "" {
 		return nil
 	}
@@ -528,7 +924,7 @@ func (runner *ContainerRunner) CaptureOutput() error {
 	_, err = os.Stat(collectionMetafile)
 	if err != nil {
 		// Regular directory
-		cw := CollectionWriter{runner.Kc, nil, sync.Mutex{}}
+		cw := CollectionWriter{0, runner.Kc, nil, nil, sync.Mutex{}}
 		manifestText, err = cw.WriteTree(runner.HostOutputDir, runner.CrunchLog.Logger)
 		if err != nil {
 			return fmt.Errorf("While uploading output files: %v", err)
@@ -541,7 +937,7 @@ func (runner *ContainerRunner) CaptureOutput() error {
 		}
 		defer file.Close()
 
-		rec := CollectionRecord{}
+		var rec arvados.Collection
 		err = json.NewDecoder(file).Decode(&rec)
 		if err != nil {
 			return fmt.Errorf("While reading FUSE metafile: %v", err)
@@ -549,20 +945,94 @@ func (runner *ContainerRunner) CaptureOutput() error {
 		manifestText = rec.ManifestText
 	}
 
-	var response CollectionRecord
+	// Pre-populate output from the configured mount points
+	var binds []string
+	for bind, _ := range runner.Container.Mounts {
+		binds = append(binds, bind)
+	}
+	sort.Strings(binds)
+
+	for _, bind := range binds {
+		mnt := runner.Container.Mounts[bind]
+
+		bindSuffix := strings.TrimPrefix(bind, runner.Container.OutputPath)
+
+		if bindSuffix == bind || len(bindSuffix) <= 0 {
+			// either does not start with OutputPath or is OutputPath itself
+			continue
+		}
+
+		if mnt.ExcludeFromOutput == true {
+			continue
+		}
+
+		// append to manifest_text
+		m, err := runner.getCollectionManifestForPath(mnt, bindSuffix)
+		if err != nil {
+			return err
+		}
+
+		manifestText = manifestText + m
+	}
+
+	// Save output
+	var response arvados.Collection
+	manifest := manifest.Manifest{Text: manifestText}
+	manifestText = manifest.Extract(".", ".").Text
 	err = runner.ArvClient.Create("collections",
 		arvadosclient.Dict{
+			"ensure_unique_name": true,
 			"collection": arvadosclient.Dict{
+				"is_trashed":    true,
+				"name":          "output for " + runner.Container.UUID,
 				"manifest_text": manifestText}},
 		&response)
 	if err != nil {
 		return fmt.Errorf("While creating output collection: %v", err)
 	}
+	runner.OutputPDH = &response.PortableDataHash
+	return nil
+}
 
-	runner.OutputPDH = new(string)
-	*runner.OutputPDH = response.PortableDataHash
+var outputCollections = make(map[string]arvados.Collection)
+
+// Fetch the collection for the mnt.PortableDataHash
+// Return the manifest_text fragment corresponding to the specified mnt.Path
+//  after making any required updates.
+//  Ex:
+//    If mnt.Path is not specified,
+//      return the entire manifest_text after replacing any "." with bindSuffix
+//    If mnt.Path corresponds to one stream,
+//      return the manifest_text for that stream after replacing that stream name with bindSuffix
+//    Otherwise, check if a filename in any one stream is being sought. Return the manifest_text
+//      for that stream after replacing stream name with bindSuffix minus the last word
+//      and the file name with last word of the bindSuffix
+//  Allowed path examples:
+//    "path":"/"
+//    "path":"/subdir1"
+//    "path":"/subdir1/subdir2"
+//    "path":"/subdir/filename" etc
+func (runner *ContainerRunner) getCollectionManifestForPath(mnt arvados.Mount, bindSuffix string) (string, error) {
+	collection := outputCollections[mnt.PortableDataHash]
+	if collection.PortableDataHash == "" {
+		err := runner.ArvClient.Get("collections", mnt.PortableDataHash, nil, &collection)
+		if err != nil {
+			return "", fmt.Errorf("While getting collection for %v: %v", mnt.PortableDataHash, err)
+		}
+		outputCollections[mnt.PortableDataHash] = collection
+	}
 
-	return nil
+	if collection.ManifestText == "" {
+		runner.CrunchLog.Printf("No manifest text for collection %v", collection.PortableDataHash)
+		return "", nil
+	}
+
+	mft := manifest.Manifest{Text: collection.ManifestText}
+	extracted := mft.Extract(mnt.Path, bindSuffix)
+	if extracted.Err != nil {
+		return "", fmt.Errorf("Error parsing manifest for %v: %v", mnt.PortableDataHash, extracted.Err.Error())
+	}
+	return extracted.Text, nil
 }
 
 func (runner *ContainerRunner) CleanupDirs() {
@@ -596,64 +1066,101 @@ func (runner *ContainerRunner) CommitLogs() error {
 	// point, but re-open crunch log with ArvClient in case there are any
 	// other further (such as failing to write the log to Keep!) while
 	// shutting down
-	runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{runner.ArvClient, runner.ContainerRecord.UUID,
-		"crunch-run", nil})
+	runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{ArvClient: runner.ArvClient,
+		UUID: runner.Container.UUID, loggingStream: "crunch-run", writeCloser: nil})
+
+	if runner.LogsPDH != nil {
+		// If we have already assigned something to LogsPDH,
+		// we must be closing the re-opened log, which won't
+		// end up getting attached to the container record and
+		// therefore doesn't need to be saved as a collection
+		// -- it exists only to send logs to other channels.
+		return nil
+	}
 
 	mt, err := runner.LogCollection.ManifestText()
 	if err != nil {
 		return fmt.Errorf("While creating log manifest: %v", err)
 	}
 
-	var response CollectionRecord
+	var response arvados.Collection
 	err = runner.ArvClient.Create("collections",
 		arvadosclient.Dict{
+			"ensure_unique_name": true,
 			"collection": arvadosclient.Dict{
-				"name":          "logs for " + runner.ContainerRecord.UUID,
+				"is_trashed":    true,
+				"name":          "logs for " + runner.Container.UUID,
 				"manifest_text": mt}},
 		&response)
 	if err != nil {
 		return fmt.Errorf("While creating log collection: %v", err)
 	}
-
-	runner.LogsPDH = new(string)
-	*runner.LogsPDH = response.PortableDataHash
-
+	runner.LogsPDH = &response.PortableDataHash
 	return nil
 }
 
-// UpdateContainerRecordRunning updates the container state to "Running"
-func (runner *ContainerRunner) UpdateContainerRecordRunning() error {
-	return runner.ArvClient.Update("containers", runner.ContainerRecord.UUID,
+// UpdateContainerRunning updates the container state to "Running"
+func (runner *ContainerRunner) UpdateContainerRunning() error {
+	runner.cStateLock.Lock()
+	defer runner.cStateLock.Unlock()
+	if runner.cCancelled {
+		return ErrCancelled
+	}
+	return runner.ArvClient.Update("containers", runner.Container.UUID,
 		arvadosclient.Dict{"container": arvadosclient.Dict{"state": "Running"}}, nil)
 }
 
-// UpdateContainerRecordComplete updates the container record state on API
+// ContainerToken returns the api_token the container (and any
+// arv-mount processes) are allowed to use.
+func (runner *ContainerRunner) ContainerToken() (string, error) {
+	if runner.token != "" {
+		return runner.token, nil
+	}
+
+	var auth arvados.APIClientAuthorization
+	err := runner.ArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
+	if err != nil {
+		return "", err
+	}
+	runner.token = auth.APIToken
+	return runner.token, nil
+}
+
+// UpdateContainerComplete updates the container record state on API
 // server to "Complete" or "Cancelled"
-func (runner *ContainerRunner) UpdateContainerRecordComplete() error {
+func (runner *ContainerRunner) UpdateContainerFinal() error {
 	update := arvadosclient.Dict{}
+	update["state"] = runner.finalState
 	if runner.LogsPDH != nil {
 		update["log"] = *runner.LogsPDH
 	}
-	if runner.ExitCode != nil {
-		update["exit_code"] = *runner.ExitCode
-	}
-	if runner.OutputPDH != nil {
-		update["output"] = runner.OutputPDH
+	if runner.finalState == "Complete" {
+		if runner.ExitCode != nil {
+			update["exit_code"] = *runner.ExitCode
+		}
+		if runner.OutputPDH != nil {
+			update["output"] = *runner.OutputPDH
+		}
 	}
+	return runner.ArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{"container": update}, nil)
+}
 
-	update["state"] = runner.finalState
-
-	return runner.ArvClient.Update("containers", runner.ContainerRecord.UUID, arvadosclient.Dict{"container": update}, nil)
+// IsCancelled returns the value of Cancelled, with goroutine safety.
+func (runner *ContainerRunner) IsCancelled() bool {
+	runner.cStateLock.Lock()
+	defer runner.cStateLock.Unlock()
+	return runner.cCancelled
 }
 
 // NewArvLogWriter creates an ArvLogWriter
 func (runner *ContainerRunner) NewArvLogWriter(name string) io.WriteCloser {
-	return &ArvLogWriter{runner.ArvClient, runner.ContainerRecord.UUID, name, runner.LogCollection.Open(name + ".txt")}
+	return &ArvLogWriter{ArvClient: runner.ArvClient, UUID: runner.Container.UUID, loggingStream: name,
+		writeCloser: runner.LogCollection.Open(name + ".txt")}
 }
 
 // Run the full container lifecycle.
 func (runner *ContainerRunner) Run() (err error) {
-	runner.CrunchLog.Printf("Executing container '%s'", runner.ContainerRecord.UUID)
+	runner.CrunchLog.Printf("Executing container '%s'", runner.Container.UUID)
 
 	hostname, hosterr := os.Hostname()
 	if hosterr != nil {
@@ -662,93 +1169,115 @@ func (runner *ContainerRunner) Run() (err error) {
 		runner.CrunchLog.Printf("Executing on host '%s'", hostname)
 	}
 
-	var runerr, waiterr error
+	// Clean up temporary directories _after_ finalizing
+	// everything (if we've made any by then)
+	defer runner.CleanupDirs()
+
+	runner.finalState = "Queued"
 
 	defer func() {
-		if err != nil {
-			runner.CrunchLog.Print(err)
+		// checkErr prints e (unless it's nil) and sets err to
+		// e (unless err is already non-nil). Thus, if err
+		// hasn't already been assigned when Run() returns,
+		// this cleanup func will cause Run() to return the
+		// first non-nil error that is passed to checkErr().
+		checkErr := func(e error) {
+			if e == nil {
+				return
+			}
+			runner.CrunchLog.Print(e)
+			if err == nil {
+				err = e
+			}
 		}
 
-		if runner.Cancelled {
-			runner.finalState = "Cancelled"
-		} else {
-			runner.finalState = "Complete"
-		}
+		// Log the error encountered in Run(), if any
+		checkErr(err)
 
-		// (6) capture output
-		outputerr := runner.CaptureOutput()
-		if outputerr != nil {
-			runner.CrunchLog.Print(outputerr)
+		if runner.finalState == "Queued" {
+			runner.CrunchLog.Close()
+			runner.UpdateContainerFinal()
+			return
 		}
 
-		// (7) clean up temporary directories
-		runner.CleanupDirs()
-
-		// (8) write logs
-		logerr := runner.CommitLogs()
-		if logerr != nil {
-			runner.CrunchLog.Print(logerr)
+		if runner.IsCancelled() {
+			runner.finalState = "Cancelled"
+			// but don't return yet -- we still want to
+			// capture partial output and write logs
 		}
 
-		// (9) update container record with results
-		updateerr := runner.UpdateContainerRecordComplete()
-		if updateerr != nil {
-			runner.CrunchLog.Print(updateerr)
-		}
+		checkErr(runner.CaptureOutput())
+		checkErr(runner.CommitLogs())
+		checkErr(runner.UpdateContainerFinal())
 
+		// The real log is already closed, but then we opened
+		// a new one in case we needed to log anything while
+		// finalizing.
 		runner.CrunchLog.Close()
-
-		if err == nil {
-			if runerr != nil {
-				err = runerr
-			} else if waiterr != nil {
-				err = waiterr
-			} else if logerr != nil {
-				err = logerr
-			} else if updateerr != nil {
-				err = updateerr
-			}
-		}
 	}()
 
-	err = runner.ArvClient.Get("containers", runner.ContainerRecord.UUID, nil, &runner.ContainerRecord)
+	err = runner.ArvClient.Get("containers", runner.Container.UUID, nil, &runner.Container)
 	if err != nil {
-		return fmt.Errorf("While getting container record: %v", err)
+		err = fmt.Errorf("While getting container record: %v", err)
+		return
 	}
 
-	// (1) setup signal handling
+	// setup signal handling
 	runner.SetupSignals()
 
-	// (2) check for and/or load image
+	// check for and/or load image
 	err = runner.LoadImage()
 	if err != nil {
-		return fmt.Errorf("While loading container image: %v", err)
+		runner.finalState = "Cancelled"
+		err = fmt.Errorf("While loading container image: %v", err)
+		return
 	}
 
-	// (3) set up FUSE mount and binds
+	// set up FUSE mount and binds
 	err = runner.SetupMounts()
 	if err != nil {
-		return fmt.Errorf("While setting up mounts: %v", err)
+		runner.finalState = "Cancelled"
+		err = fmt.Errorf("While setting up mounts: %v", err)
+		return
+	}
+
+	err = runner.CreateContainer()
+	if err != nil {
+		return
 	}
 
-	// (3) create and start container
-	err = runner.StartContainer()
+	// Gather and record node information
+	err = runner.LogNodeInfo()
 	if err != nil {
-		if err == ErrCancelled {
-			err = nil
-		}
+		return
+	}
+	// Save container.json record on log collection
+	err = runner.LogContainerRecord()
+	if err != nil {
+		return
+	}
+
+	runner.StartCrunchstat()
+
+	if runner.IsCancelled() {
 		return
 	}
 
-	// (4) update container record state
-	err = runner.UpdateContainerRecordRunning()
+	err = runner.UpdateContainerRunning()
 	if err != nil {
-		runner.CrunchLog.Print(err)
+		return
 	}
+	runner.finalState = "Cancelled"
 
-	// (5) wait for container to finish
-	waiterr = runner.WaitFinish()
+	err = runner.StartContainer()
+	if err != nil {
+		return
+	}
 
+	err = runner.WaitFinish()
+	if err == nil {
+		runner.finalState = "Complete"
+	}
 	return
 }
 
@@ -762,18 +1291,38 @@ func NewContainerRunner(api IArvadosClient,
 	cr.NewLogWriter = cr.NewArvLogWriter
 	cr.RunArvMount = cr.ArvMountCmd
 	cr.MkTempDir = ioutil.TempDir
-	cr.LogCollection = &CollectionWriter{kc, nil, sync.Mutex{}}
-	cr.ContainerRecord.UUID = containerUUID
+	cr.LogCollection = &CollectionWriter{0, kc, nil, nil, sync.Mutex{}}
+	cr.Container.UUID = containerUUID
 	cr.CrunchLog = NewThrottledLogger(cr.NewLogWriter("crunch-run"))
 	cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
+
+	loadLogThrottleParams(api)
+
 	return cr
 }
 
 func main() {
+	statInterval := flag.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
+	cgroupRoot := flag.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree")
+	cgroupParent := flag.String("cgroup-parent", "docker", "name of container's parent cgroup (ignored if -cgroup-parent-subsystem is used)")
+	cgroupParentSubsystem := flag.String("cgroup-parent-subsystem", "", "use current cgroup for given subsystem as parent cgroup for container")
+	caCertsPath := flag.String("ca-certs", "", "Path to TLS root certificates")
+	enableNetwork := flag.String("container-enable-networking", "default",
+		`Specify if networking should be enabled for container.  One of 'default', 'always':
+    	default: only enable networking if container requests it.
+    	always:  containers always have networking enabled
+    	`)
+	networkMode := flag.String("container-network-mode", "default",
+		`Set networking mode for container.  Corresponds to Docker network mode (--net).
+    	`)
 	flag.Parse()
 
 	containerId := flag.Arg(0)
 
+	if *caCertsPath != "" {
+		arvadosclient.CertFiles = []string{*caCertsPath}
+	}
+
 	api, err := arvadosclient.MakeArvadosClient()
 	if err != nil {
 		log.Fatalf("%s: %v", containerId, err)
@@ -781,19 +1330,33 @@ func main() {
 	api.Retries = 8
 
 	var kc *keepclient.KeepClient
-	kc, err = keepclient.MakeKeepClient(&api)
+	kc, err = keepclient.MakeKeepClient(api)
 	if err != nil {
 		log.Fatalf("%s: %v", containerId, err)
 	}
 	kc.Retries = 4
 
-	var docker *dockerclient.DockerClient
-	docker, err = dockerclient.NewDockerClient("unix:///var/run/docker.sock", nil)
+	var docker *dockerclient.Client
+	// API version 1.21 corresponds to Docker 1.9, which is currently the
+	// minimum version we want to support.
+	docker, err = dockerclient.NewClient(dockerclient.DefaultDockerHost, "1.21", nil, nil)
 	if err != nil {
 		log.Fatalf("%s: %v", containerId, err)
 	}
 
-	cr := NewContainerRunner(api, kc, docker, containerId)
+	dockerClientProxy := ThinDockerClientProxy{Docker: docker}
+
+	cr := NewContainerRunner(api, kc, dockerClientProxy, containerId)
+	cr.statInterval = *statInterval
+	cr.cgroupRoot = *cgroupRoot
+	cr.expectCgroupParent = *cgroupParent
+	cr.enableNetwork = *enableNetwork
+	cr.networkMode = *networkMode
+	if *cgroupParentSubsystem != "" {
+		p := findCgroup(*cgroupParentSubsystem)
+		cr.setCgroupParent = p
+		cr.expectCgroupParent = p
+	}
 
 	err = cr.Run()
 	if err != nil {