X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/d18557927ac033c2e33c1f7770881367b252f329..a19ffad966b25b3869e666f749f7c6da187bef68:/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go

diff --git a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go
index ef08f4583a..60dc6071b4 100644
--- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go
+++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go
@@ -3,15 +3,14 @@ package main
 // Dispatcher service for Crunch that submits containers to the slurm queue.
 
 import (
-	"encoding/json"
+	"bytes"
 	"flag"
 	"fmt"
 	"git.curoverse.com/arvados.git/sdk/go/arvados"
 	"git.curoverse.com/arvados.git/sdk/go/arvadosclient"
+	"git.curoverse.com/arvados.git/sdk/go/config"
 	"git.curoverse.com/arvados.git/sdk/go/dispatch"
 	"github.com/coreos/go-systemd/daemon"
-	"io"
-	"io/ioutil"
 	"log"
 	"math"
 	"os"
@@ -32,21 +31,24 @@ type Config struct {
 	//
 	// Example: []string{"crunch-run", "--cgroup-parent-subsystem=memory"}
 	CrunchRunCommand []string
+
+	// Minimum time between two attempts to run the same container
+	MinRetryPeriod arvados.Duration
 }
 
 func main() {
 	err := doMain()
 	if err != nil {
-		log.Fatalf("%q", err)
+		log.Fatal(err)
 	}
 }
 
 var (
-	config        Config
-	squeueUpdater Squeue
+	theConfig Config
+	sqCheck   SqueueChecker
 )
 
-const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/config.json"
+const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
 
 func doMain() error {
 	flags := flag.NewFlagSet("crunch-dispatch-slurm", flag.ExitOnError)
@@ -55,36 +57,35 @@ func doMain() error {
 	configPath := flags.String(
 		"config",
 		defaultConfigPath,
-		"`path` to json configuration file")
+		"`path` to JSON or YAML configuration file")
 
 	// Parse args; omit the first arg which is the command name
 	flags.Parse(os.Args[1:])
 
-	err := readConfig(&config, *configPath)
+	err := readConfig(&theConfig, *configPath)
 	if err != nil {
-		log.Printf("Error reading configuration: %v", err)
 		return err
 	}
 
-	if config.CrunchRunCommand == nil {
-		config.CrunchRunCommand = []string{"crunch-run"}
+	if theConfig.CrunchRunCommand == nil {
+		theConfig.CrunchRunCommand = []string{"crunch-run"}
 	}
 
-	if config.PollPeriod == 0 {
-		config.PollPeriod = arvados.Duration(10 * time.Second)
+	if theConfig.PollPeriod == 0 {
+		theConfig.PollPeriod = arvados.Duration(10 * time.Second)
 	}
 
-	if config.Client.APIHost != "" || config.Client.AuthToken != "" {
+	if theConfig.Client.APIHost != "" || theConfig.Client.AuthToken != "" {
 		// Copy real configs into env vars so [a]
 		// MakeArvadosClient() uses them, and [b] they get
 		// propagated to crunch-run via SLURM.
-		os.Setenv("ARVADOS_API_HOST", config.Client.APIHost)
-		os.Setenv("ARVADOS_API_TOKEN", config.Client.AuthToken)
-		os.Setenv("ARVADOS_API_INSECURE", "")
-		if config.Client.Insecure {
-			os.Setenv("ARVADOS_API_INSECURE", "1")
+		os.Setenv("ARVADOS_API_HOST", theConfig.Client.APIHost)
+		os.Setenv("ARVADOS_API_TOKEN", theConfig.Client.AuthToken)
+		os.Setenv("ARVADOS_API_HOST_INSECURE", "")
+		if theConfig.Client.Insecure {
+			os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
 		}
-		os.Setenv("ARVADOS_KEEP_SERVICES", "")
+		os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(theConfig.Client.KeepServiceURIs, " "))
 		os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
 	} else {
 		log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).")
@@ -97,25 +98,21 @@ func doMain() error {
 	}
 	arv.Retries = 25
 
-	squeueUpdater.StartMonitor(time.Duration(config.PollPeriod))
-	defer squeueUpdater.Done()
+	sqCheck = SqueueChecker{Period: time.Duration(theConfig.PollPeriod)}
+	defer sqCheck.Stop()
 
 	dispatcher := dispatch.Dispatcher{
 		Arv:            arv,
 		RunContainer:   run,
-		PollInterval:   time.Duration(config.PollPeriod),
-		DoneProcessing: make(chan struct{})}
-
-	if _, err := daemon.SdNotify("READY=1"); err != nil {
-		log.Printf("Error notifying init daemon: %v", err)
+		PollPeriod:     time.Duration(theConfig.PollPeriod),
+		MinRetryPeriod: time.Duration(theConfig.MinRetryPeriod),
 	}
 
-	err = dispatcher.RunDispatcher()
-	if err != nil {
-		return err
+	if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
+		log.Printf("Error notifying init daemon: %v", err)
 	}
 
-	return nil
+	return dispatcher.Run()
 }
 
 // sbatchCmd
@@ -124,10 +121,13 @@ func sbatchFunc(container arvados.Container) *exec.Cmd {
 
 	var sbatchArgs []string
 	sbatchArgs = append(sbatchArgs, "--share")
-	sbatchArgs = append(sbatchArgs, config.SbatchArguments...)
+	sbatchArgs = append(sbatchArgs, theConfig.SbatchArguments...)
 	sbatchArgs = append(sbatchArgs, fmt.Sprintf("--job-name=%s", container.UUID))
 	sbatchArgs = append(sbatchArgs, fmt.Sprintf("--mem-per-cpu=%d", int(memPerCPU)))
 	sbatchArgs = append(sbatchArgs, fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs))
+	if container.SchedulingParameters.Partitions != nil {
+		sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ",")))
+	}
 
 	return exec.Command("sbatch", sbatchArgs...)
 }
@@ -151,79 +151,37 @@ func submit(dispatcher *dispatch.Dispatcher,
 			// OK, no cleanup needed
 			return
 		}
-		err := dispatcher.Arv.Update("containers", container.UUID,
-			arvadosclient.Dict{
-				"container": arvadosclient.Dict{"state": "Queued"}},
-			nil)
+		err := dispatcher.Unlock(container.UUID)
 		if err != nil {
 			log.Printf("Error unlocking container %s: %v", container.UUID, err)
 		}
 	}()
 
-	// Create the command and attach to stdin/stdout
 	cmd := sbatchCmd(container)
-	stdinWriter, stdinerr := cmd.StdinPipe()
-	if stdinerr != nil {
-		submitErr = fmt.Errorf("Error creating stdin pipe %v: %q", container.UUID, stdinerr)
-		return
-	}
 
-	stdoutReader, stdoutErr := cmd.StdoutPipe()
-	if stdoutErr != nil {
-		submitErr = fmt.Errorf("Error creating stdout pipe %v: %q", container.UUID, stdoutErr)
-		return
-	}
+	// Send a tiny script on stdin to execute the crunch-run
+	// command (slurm requires this to be a #! script)
+	cmd.Stdin = strings.NewReader(execScript(append(crunchRunCommand, container.UUID)))
 
-	stderrReader, stderrErr := cmd.StderrPipe()
-	if stderrErr != nil {
-		submitErr = fmt.Errorf("Error creating stderr pipe %v: %q", container.UUID, stderrErr)
-		return
-	}
+	var stdout, stderr bytes.Buffer
+	cmd.Stdout = &stdout
+	cmd.Stderr = &stderr
 
 	// Mutex between squeue sync and running sbatch or scancel.
-	squeueUpdater.SlurmLock.Lock()
-	defer squeueUpdater.SlurmLock.Unlock()
-
-	err := cmd.Start()
-	if err != nil {
-		submitErr = fmt.Errorf("Error starting %v: %v", cmd.Args, err)
-		return
+	sqCheck.L.Lock()
+	defer sqCheck.L.Unlock()
+
+	log.Printf("exec sbatch %+q", cmd.Args)
+	err := cmd.Run()
+	switch err.(type) {
+	case nil:
+		log.Printf("sbatch succeeded: %q", strings.TrimSpace(stdout.String()))
+		return nil
+	case *exec.ExitError:
+		return fmt.Errorf("sbatch %+q failed: %v (stderr: %q)", cmd.Args, err, stderr)
+	default:
+		return fmt.Errorf("exec failed: %v", err)
 	}
-
-	stdoutChan := make(chan []byte)
-	go func() {
-		b, _ := ioutil.ReadAll(stdoutReader)
-		stdoutReader.Close()
-		stdoutChan <- b
-	}()
-
-	stderrChan := make(chan []byte)
-	go func() {
-		b, _ := ioutil.ReadAll(stderrReader)
-		stderrReader.Close()
-		stderrChan <- b
-	}()
-
-	// Send a tiny script on stdin to execute the crunch-run command
-	// slurm actually enforces that this must be a #! script
-	io.WriteString(stdinWriter, execScript(append(crunchRunCommand, container.UUID)))
-	stdinWriter.Close()
-
-	err = cmd.Wait()
-
-	stdoutMsg := <-stdoutChan
-	stderrmsg := <-stderrChan
-
-	close(stdoutChan)
-	close(stderrChan)
-
-	if err != nil {
-		submitErr = fmt.Errorf("Container submission failed: %v: %v (stderr: %q)", cmd.Args, err, stderrmsg)
-		return
-	}
-
-	log.Printf("sbatch succeeded: %s", strings.TrimSpace(string(stdoutMsg)))
-	return
 }
 
 // If the container is marked as Locked, check if it is already in the slurm
@@ -234,7 +192,7 @@ func submit(dispatcher *dispatch.Dispatcher,
 func monitorSubmitOrCancel(dispatcher *dispatch.Dispatcher, container arvados.Container, monitorDone *bool) {
 	submitted := false
 	for !*monitorDone {
-		if squeueUpdater.CheckSqueue(container.UUID) {
+		if sqCheck.HasUUID(container.UUID) {
 			// Found in the queue, so continue monitoring
 			submitted = true
 		} else if container.State == dispatch.Locked && !submitted {
@@ -243,11 +201,11 @@ func monitorSubmitOrCancel(dispatcher *dispatch.Dispatcher, container arvados.Co
 
 			log.Printf("About to submit queued container %v", container.UUID)
 
-			if err := submit(dispatcher, container, config.CrunchRunCommand); err != nil {
+			if err := submit(dispatcher, container, theConfig.CrunchRunCommand); err != nil {
 				log.Printf("Error submitting container %s to slurm: %v",
 					container.UUID, err)
 				// maybe sbatch is broken, put it back to queued
-				dispatcher.UpdateState(container.UUID, dispatch.Queued)
+				dispatcher.Unlock(container.UUID)
 			}
 			submitted = true
 		} else {
@@ -263,20 +221,20 @@ func monitorSubmitOrCancel(dispatcher *dispatch.Dispatcher, container arvados.Co
 				log.Printf("Error getting final container state: %v", err)
 			}
 
-			var st arvados.ContainerState
 			switch con.State {
 			case dispatch.Locked:
-				st = dispatch.Queued
+				log.Printf("Container %s in state %v but missing from slurm queue, changing to %v.",
+					container.UUID, con.State, dispatch.Queued)
+				dispatcher.Unlock(container.UUID)
 			case dispatch.Running:
-				st = dispatch.Cancelled
+				st := dispatch.Cancelled
+				log.Printf("Container %s in state %v but missing from slurm queue, changing to %v.",
+					container.UUID, con.State, st)
+				dispatcher.UpdateState(container.UUID, st)
 			default:
 				// Container state is Queued, Complete or Cancelled so stop monitoring it.
 				return
 			}
-
-			log.Printf("Container %s in state %v but missing from slurm queue, changing to %v.",
-				container.UUID, con.State, st)
-			dispatcher.UpdateState(container.UUID, st)
 		}
 	}
 }
@@ -296,43 +254,34 @@ func run(dispatcher *dispatch.Dispatcher,
 	go monitorSubmitOrCancel(dispatcher, container, &monitorDone)
 
 	for container = range status {
-		if container.State == dispatch.Locked || container.State == dispatch.Running {
-			if container.Priority == 0 {
-				log.Printf("Canceling container %s", container.UUID)
-
-				// Mutex between squeue sync and running sbatch or scancel.
-				squeueUpdater.SlurmLock.Lock()
-				err := scancelCmd(container).Run()
-				squeueUpdater.SlurmLock.Unlock()
-
-				if err != nil {
-					log.Printf("Error stopping container %s with scancel: %v",
-						container.UUID, err)
-					if squeueUpdater.CheckSqueue(container.UUID) {
-						log.Printf("Container %s is still in squeue after scancel.",
-							container.UUID)
-						continue
-					}
-				}
+		if container.Priority == 0 && (container.State == dispatch.Locked || container.State == dispatch.Running) {
+			log.Printf("Canceling container %s", container.UUID)
+			// Mutex between squeue sync and running sbatch or scancel.
+			sqCheck.L.Lock()
+			cmd := scancelCmd(container)
+			msg, err := cmd.CombinedOutput()
+			sqCheck.L.Unlock()
 
-				err = dispatcher.UpdateState(container.UUID, dispatch.Cancelled)
+			if err != nil {
+				log.Printf("Error stopping container %s with %v %v: %v %v", container.UUID, cmd.Path, cmd.Args, err, string(msg))
+				if sqCheck.HasUUID(container.UUID) {
+					log.Printf("Container %s is still in squeue after scancel.", container.UUID)
+					continue
+				}
 			}
+
+			// Ignore errors; if necessary, we'll try again next time
+			dispatcher.UpdateState(container.UUID, dispatch.Cancelled)
 		}
 	}
 	monitorDone = true
 }
 
 func readConfig(dst interface{}, path string) error {
-	if buf, err := ioutil.ReadFile(path); err != nil && os.IsNotExist(err) {
-		if path == defaultConfigPath {
-			log.Printf("Config not specified. Continue with default configuration.")
-		} else {
-			return fmt.Errorf("Config file not found %q: %v", path, err)
-		}
-	} else if err != nil {
-		return fmt.Errorf("Error reading config %q: %v", path, err)
-	} else if err = json.Unmarshal(buf, dst); err != nil {
-		return fmt.Errorf("Error decoding config %q: %v", path, err)
+	err := config.LoadFile(dst, path)
+	if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
+		log.Printf("Config not specified. Continue with default configuration.")
+		err = nil
 	}
-	return nil
+	return err
 }