X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/ae5eb12d3d9ab298a4c36412b4a4d83272574d25..10f08c358c12468119dc2621c48b68d6d33417da:/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 8cf2c00e9e..879cb785d9 100644
--- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go
+++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go
@@ -1,26 +1,38 @@
+// Copyright (C) The Arvados Authors. All rights reserved.
+//
+// SPDX-License-Identifier: AGPL-3.0
+
 package main
 
 // Dispatcher service for Crunch that submits containers to the slurm queue.
 
 import (
-	"bytes"
+	"context"
 	"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"
 	"log"
 	"math"
 	"os"
-	"os/exec"
+	"regexp"
 	"strings"
 	"time"
+
+	"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"
+	"git.curoverse.com/arvados.git/services/dispatchcloud"
+	"github.com/coreos/go-systemd/daemon"
 )
 
-// Config used by crunch-dispatch-slurm
-type Config struct {
+var version = "dev"
+
+type command struct {
+	dispatcher *dispatch.Dispatcher
+	cluster    *arvados.Cluster
+	sqCheck    *SqueueChecker
+	slurm      Slurm
+
 	Client arvados.Client
 
 	SbatchArguments []string
@@ -37,60 +49,74 @@ type Config struct {
 }
 
 func main() {
-	err := doMain()
+	err := (&command{}).Run(os.Args[0], os.Args[1:])
 	if err != nil {
 		log.Fatal(err)
 	}
 }
 
-var (
-	theConfig     Config
-	squeueUpdater Squeue
-)
-
 const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
 
-func doMain() error {
-	flags := flag.NewFlagSet("crunch-dispatch-slurm", flag.ExitOnError)
+func (cmd *command) Run(prog string, args []string) error {
+	flags := flag.NewFlagSet(prog, flag.ExitOnError)
 	flags.Usage = func() { usage(flags) }
 
 	configPath := flags.String(
 		"config",
 		defaultConfigPath,
 		"`path` to JSON or YAML configuration file")
-
+	dumpConfig := flag.Bool(
+		"dump-config",
+		false,
+		"write current configuration to stdout and exit")
+	getVersion := flags.Bool(
+		"version",
+		false,
+		"Print version information and exit.")
 	// Parse args; omit the first arg which is the command name
-	flags.Parse(os.Args[1:])
+	flags.Parse(args)
+
+	// Print version information if requested
+	if *getVersion {
+		fmt.Printf("crunch-dispatch-slurm %s\n", version)
+		return nil
+	}
+
+	log.Printf("crunch-dispatch-slurm %s started", version)
 
-	err := readConfig(&theConfig, *configPath)
+	err := cmd.readConfig(*configPath)
 	if err != nil {
 		return err
 	}
 
-	if theConfig.CrunchRunCommand == nil {
-		theConfig.CrunchRunCommand = []string{"crunch-run"}
+	if cmd.CrunchRunCommand == nil {
+		cmd.CrunchRunCommand = []string{"crunch-run"}
 	}
 
-	if theConfig.PollPeriod == 0 {
-		theConfig.PollPeriod = arvados.Duration(10 * time.Second)
+	if cmd.PollPeriod == 0 {
+		cmd.PollPeriod = arvados.Duration(10 * time.Second)
 	}
 
-	if theConfig.Client.APIHost != "" || theConfig.Client.AuthToken != "" {
+	if cmd.Client.APIHost != "" || cmd.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", theConfig.Client.APIHost)
-		os.Setenv("ARVADOS_API_TOKEN", theConfig.Client.AuthToken)
+		os.Setenv("ARVADOS_API_HOST", cmd.Client.APIHost)
+		os.Setenv("ARVADOS_API_TOKEN", cmd.Client.AuthToken)
 		os.Setenv("ARVADOS_API_HOST_INSECURE", "")
-		if theConfig.Client.Insecure {
+		if cmd.Client.Insecure {
 			os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
 		}
-		os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(theConfig.Client.KeepServiceURIs, " "))
+		os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(cmd.Client.KeepServiceURIs, " "))
 		os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
 	} else {
 		log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).")
 	}
 
+	if *dumpConfig {
+		log.Fatal(config.DumpAndExit(cmd))
+	}
+
 	arv, err := arvadosclient.MakeArvadosClient()
 	if err != nil {
 		log.Printf("Error making Arvados client: %v", err)
@@ -98,192 +124,235 @@ func doMain() error {
 	}
 	arv.Retries = 25
 
-	squeueUpdater.StartMonitor(time.Duration(theConfig.PollPeriod))
-	defer squeueUpdater.Done()
+	siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile)
+	if os.IsNotExist(err) {
+		log.Printf("warning: no cluster config file %q (%s), proceeding with no node types defined", arvados.DefaultConfigFile, err)
+	} else if err != nil {
+		log.Fatalf("error loading config: %s", err)
+	} else if cmd.cluster, err = siteConfig.GetCluster(""); err != nil {
+		log.Fatalf("config error: %s", err)
+	} else if len(cmd.cluster.InstanceTypes) > 0 {
+		go dispatchcloud.SlurmNodeTypeFeatureKludge(cmd.cluster)
+	}
+
+	if cmd.slurm == nil {
+		cmd.slurm = &slurmCLI{}
+	}
 
-	dispatcher := dispatch.Dispatcher{
+	cmd.sqCheck = &SqueueChecker{
+		Period: time.Duration(cmd.PollPeriod),
+		Slurm:  cmd.slurm,
+	}
+	defer cmd.sqCheck.Stop()
+
+	cmd.dispatcher = &dispatch.Dispatcher{
 		Arv:            arv,
-		RunContainer:   run,
-		PollInterval:   time.Duration(theConfig.PollPeriod),
-		MinRetryPeriod: time.Duration(theConfig.MinRetryPeriod),
+		RunContainer:   cmd.run,
+		PollPeriod:     time.Duration(cmd.PollPeriod),
+		MinRetryPeriod: time.Duration(cmd.MinRetryPeriod),
 	}
 
 	if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
 		log.Printf("Error notifying init daemon: %v", err)
 	}
 
-	return dispatcher.Run()
+	go cmd.checkSqueueForOrphans()
+
+	return cmd.dispatcher.Run(context.Background())
+}
+
+var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$`)
+
+// Check the next squeue report, and invoke TrackContainer for all the
+// containers in the report. This gives us a chance to cancel slurm
+// jobs started by a previous dispatch process that never released
+// their slurm allocations even though their container states are
+// Cancelled or Complete. See https://dev.arvados.org/issues/10979
+func (cmd *command) checkSqueueForOrphans() {
+	for _, uuid := range cmd.sqCheck.All() {
+		if !containerUuidPattern.MatchString(uuid) {
+			continue
+		}
+		err := cmd.dispatcher.TrackContainer(uuid)
+		if err != nil {
+			log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err)
+		}
+	}
+}
+
+func (cmd *command) niceness(priority int) int {
+	if priority > 1000 {
+		priority = 1000
+	}
+	if priority < 0 {
+		priority = 0
+	}
+	// Niceness range 1-10000
+	return (1000 - priority) * 10
 }
 
-// sbatchCmd
-func sbatchFunc(container arvados.Container) *exec.Cmd {
-	memPerCPU := math.Ceil(float64(container.RuntimeConstraints.RAM) / (float64(container.RuntimeConstraints.VCPUs) * 1048576))
+func (cmd *command) sbatchArgs(container arvados.Container) ([]string, error) {
+	mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM) / float64(1048576)))
+
+	var disk int64
+	for _, m := range container.Mounts {
+		if m.Kind == "tmp" {
+			disk += m.Capacity
+		}
+	}
+	disk = int64(math.Ceil(float64(disk) / float64(1048576)))
 
 	var sbatchArgs []string
-	sbatchArgs = append(sbatchArgs, "--share")
-	sbatchArgs = append(sbatchArgs, theConfig.SbatchArguments...)
+	sbatchArgs = append(sbatchArgs, cmd.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("--mem=%d", mem))
 	sbatchArgs = append(sbatchArgs, fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs))
-	if container.SchedulingParameters.Partitions != nil {
+	sbatchArgs = append(sbatchArgs, fmt.Sprintf("--tmp=%d", disk))
+	sbatchArgs = append(sbatchArgs, fmt.Sprintf("--nice=%d", cmd.niceness(container.Priority)))
+	if len(container.SchedulingParameters.Partitions) > 0 {
 		sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ",")))
 	}
 
-	return exec.Command("sbatch", sbatchArgs...)
-}
+	if cmd.cluster == nil {
+		// no instance types configured
+	} else if it, err := dispatchcloud.ChooseInstanceType(cmd.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured {
+		// ditto
+	} else if err != nil {
+		return nil, err
+	} else {
+		sbatchArgs = append(sbatchArgs, "--constraint="+it.Name)
+	}
 
-// scancelCmd
-func scancelFunc(container arvados.Container) *exec.Cmd {
-	return exec.Command("scancel", "--name="+container.UUID)
+	return sbatchArgs, nil
 }
 
-// Wrap these so that they can be overridden by tests
-var sbatchCmd = sbatchFunc
-var scancelCmd = scancelFunc
-
-// Submit job to slurm using sbatch.
-func submit(dispatcher *dispatch.Dispatcher,
-	container arvados.Container, crunchRunCommand []string) (submitErr error) {
-	defer func() {
-		// If we didn't get as far as submitting a slurm job,
-		// unlock the container and return it to the queue.
-		if submitErr == nil {
-			// OK, no cleanup needed
-			return
-		}
-		err := dispatcher.Unlock(container.UUID)
-		if err != nil {
-			log.Printf("Error unlocking container %s: %v", container.UUID, err)
-		}
-	}()
+func (cmd *command) submit(container arvados.Container, crunchRunCommand []string) error {
+	// append() here avoids modifying crunchRunCommand's
+	// underlying array, which is shared with other goroutines.
+	crArgs := append([]string(nil), crunchRunCommand...)
+	crArgs = append(crArgs, container.UUID)
+	crScript := strings.NewReader(execScript(crArgs))
 
-	cmd := sbatchCmd(container)
+	cmd.sqCheck.L.Lock()
+	defer cmd.sqCheck.L.Unlock()
 
-	// 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)))
+	sbArgs, err := cmd.sbatchArgs(container)
+	if err != nil {
+		return err
+	}
+	log.Printf("running sbatch %+q", sbArgs)
+	return cmd.slurm.Batch(crScript, sbArgs)
+}
 
-	var stdout, stderr bytes.Buffer
-	cmd.Stdout = &stdout
-	cmd.Stderr = &stderr
+// Submit a container to the slurm queue (or resume monitoring if it's
+// already in the queue).  Cancel the slurm job if the container's
+// priority changes to zero or its state indicates it's no longer
+// running.
+func (cmd *command) run(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	if ctr.State == dispatch.Locked && !cmd.sqCheck.HasUUID(ctr.UUID) {
+		log.Printf("Submitting container %s to slurm", ctr.UUID)
+		if err := cmd.submit(ctr, cmd.CrunchRunCommand); err != nil {
+			var text string
+			if err == dispatchcloud.ErrConstraintsNotSatisfiable {
+				text = fmt.Sprintf("cannot run container %s: %s", ctr.UUID, err)
+				cmd.dispatcher.UpdateState(ctr.UUID, dispatch.Cancelled)
+			} else {
+				text = fmt.Sprintf("Error submitting container %s to slurm: %s", ctr.UUID, err)
+			}
+			log.Print(text)
 
-	// Mutex between squeue sync and running sbatch or scancel.
-	squeueUpdater.SlurmLock.Lock()
-	defer squeueUpdater.SlurmLock.Unlock()
+			lr := arvadosclient.Dict{"log": arvadosclient.Dict{
+				"object_uuid": ctr.UUID,
+				"event_type":  "dispatch",
+				"properties":  map[string]string{"text": text}}}
+			cmd.dispatcher.Arv.Create("logs", lr, nil)
 
-	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)
+			cmd.dispatcher.Unlock(ctr.UUID)
+			return
+		}
 	}
-}
 
-// If the container is marked as Locked, check if it is already in the slurm
-// queue.  If not, submit it.
-//
-// If the container is marked as Running, check if it is in the slurm queue.
-// If not, mark it as Cancelled.
-func monitorSubmitOrCancel(dispatcher *dispatch.Dispatcher, container arvados.Container, monitorDone *bool) {
-	submitted := false
-	for !*monitorDone {
-		if squeueUpdater.CheckSqueue(container.UUID) {
-			// Found in the queue, so continue monitoring
-			submitted = true
-		} else if container.State == dispatch.Locked && !submitted {
-			// Not in queue but in Locked state and we haven't
-			// submitted it yet, so submit it.
-
-			log.Printf("About to submit queued container %v", container.UUID)
-
-			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.Unlock(container.UUID)
-			}
-			submitted = true
-		} else {
-			// Not in queue and we are not going to submit it.
-			// Refresh the container state. If it is
-			// Complete/Cancelled, do nothing, if it is Locked then
-			// release it back to the Queue, if it is Running then
-			// clean up the record.
-
-			var con arvados.Container
-			err := dispatcher.Arv.Get("containers", container.UUID, nil, &con)
-			if err != nil {
-				log.Printf("Error getting final container state: %v", err)
-			}
+	log.Printf("Start monitoring container %v in state %q", ctr.UUID, ctr.State)
+	defer log.Printf("Done monitoring container %s", ctr.UUID)
 
-			switch con.State {
-			case dispatch.Locked:
-				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)
+	// If the container disappears from the slurm queue, there is
+	// no point in waiting for further dispatch updates: just
+	// clean up and return.
+	go func(uuid string) {
+		for ctx.Err() == nil && cmd.sqCheck.HasUUID(uuid) {
+		}
+		cancel()
+	}(ctr.UUID)
+
+	for {
+		select {
+		case <-ctx.Done():
+			// Disappeared from squeue
+			if err := cmd.dispatcher.Arv.Get("containers", ctr.UUID, nil, &ctr); err != nil {
+				log.Printf("Error getting final container state for %s: %s", ctr.UUID, err)
+			}
+			switch ctr.State {
 			case dispatch.Running:
-				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
+				cmd.dispatcher.UpdateState(ctr.UUID, dispatch.Cancelled)
+			case dispatch.Locked:
+				cmd.dispatcher.Unlock(ctr.UUID)
+			}
+			return
+		case updated, ok := <-status:
+			if !ok {
+				log.Printf("Dispatcher says container %s is done: cancel slurm job", ctr.UUID)
+				cmd.scancel(ctr)
+			} else if updated.Priority == 0 {
+				log.Printf("Container %s has state %q, priority %d: cancel slurm job", ctr.UUID, updated.State, updated.Priority)
+				cmd.scancel(ctr)
+			} else {
+				cmd.renice(updated)
 			}
 		}
 	}
 }
 
-// Run or monitor a container.
-//
-// Monitor status updates.  If the priority changes to zero, cancel the
-// container using scancel.
-func run(dispatcher *dispatch.Dispatcher,
-	container arvados.Container,
-	status chan arvados.Container) {
-
-	log.Printf("Monitoring container %v started", container.UUID)
-	defer log.Printf("Monitoring container %v finished", container.UUID)
+func (cmd *command) scancel(ctr arvados.Container) {
+	cmd.sqCheck.L.Lock()
+	err := cmd.slurm.Cancel(ctr.UUID)
+	cmd.sqCheck.L.Unlock()
 
-	monitorDone := false
-	go monitorSubmitOrCancel(dispatcher, container, &monitorDone)
-
-	for container = range status {
-		if !(container.State == dispatch.Locked || container.State == dispatch.Running) {
-			continue
-		}
-		if container.Priority != 0 {
-			continue
-		}
-		log.Printf("Canceling container %s", container.UUID)
-
-		// Mutex between squeue sync and running sbatch or scancel.
-		squeueUpdater.SlurmLock.Lock()
-		cmd := scancelCmd(container)
-		msg, err := cmd.CombinedOutput()
-		squeueUpdater.SlurmLock.Unlock()
+	if err != nil {
+		log.Printf("scancel: %s", err)
+		time.Sleep(time.Second)
+	} else if cmd.sqCheck.HasUUID(ctr.UUID) {
+		log.Printf("container %s is still in squeue after scancel", ctr.UUID)
+		time.Sleep(time.Second)
+	}
+}
 
-		if err != nil {
-			log.Printf("Error stopping container %s with %v %v: %v %v", container.UUID, cmd.Path, cmd.Args, err, string(msg))
-			if squeueUpdater.CheckSqueue(container.UUID) {
-				log.Printf("Container %s is still in squeue after scancel.", container.UUID)
-				continue
-			}
-		}
+func (cmd *command) renice(ctr arvados.Container) {
+	nice := cmd.niceness(ctr.Priority)
+	oldnice := cmd.sqCheck.GetNiceness(ctr.UUID)
+	if nice == oldnice || oldnice == -1 {
+		return
+	}
+	log.Printf("updating slurm nice value to %d (was %d)", nice, oldnice)
+	cmd.sqCheck.L.Lock()
+	err := cmd.slurm.Renice(ctr.UUID, nice)
+	cmd.sqCheck.L.Unlock()
 
-		// Ignore errors; if necessary, we'll try again next time
-		dispatcher.UpdateState(container.UUID, dispatch.Cancelled)
+	if err != nil {
+		log.Printf("renice: %s", err)
+		time.Sleep(time.Second)
+		return
+	}
+	if cmd.sqCheck.HasUUID(ctr.UUID) {
+		log.Printf("container %s has arvados priority %d, slurm nice %d",
+			ctr.UUID, ctr.Priority, cmd.sqCheck.GetNiceness(ctr.UUID))
 	}
-	monitorDone = true
 }
 
-func readConfig(dst interface{}, path string) error {
-	err := config.LoadFile(dst, path)
+func (cmd *command) readConfig(path string) error {
+	err := config.LoadFile(cmd, path)
 	if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
 		log.Printf("Config not specified. Continue with default configuration.")
 		err = nil