+// 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"
"log"
"math"
"os"
- "os/exec"
+ "regexp"
"strings"
"time"
+ "git.curoverse.com/arvados.git/lib/dispatchcloud"
"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"
+ "github.com/sirupsen/logrus"
+)
+
+type logger interface {
+ dispatch.Logger
+ Fatalf(string, ...interface{})
+}
+
+const initialNiceValue int64 = 10000
+
+var (
+ version = "dev"
+ defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
)
-// Config used by crunch-dispatch-slurm
-type Config struct {
+type Dispatcher struct {
+ *dispatch.Dispatcher
+ logger logrus.FieldLogger
+ cluster *arvados.Cluster
+ sqCheck *SqueueChecker
+ slurm Slurm
+
Client arvados.Client
SbatchArguments []string
PollPeriod arvados.Duration
+ PrioritySpread int64
// crunch-run command to invoke. The container UUID will be
// appended. If nil, []string{"crunch-run"} will be used.
// Example: []string{"crunch-run", "--cgroup-parent-subsystem=memory"}
CrunchRunCommand []string
+ // Extra RAM to reserve (in Bytes) for SLURM job, in addition
+ // to the amount specified in the container's RuntimeConstraints
+ ReserveExtraRAM int64
+
// Minimum time between two attempts to run the same container
MinRetryPeriod arvados.Duration
+
+ // Batch size for container queries
+ BatchSize int64
}
func main() {
- err := doMain()
+ logger := logrus.StandardLogger()
+ if os.Getenv("DEBUG") != "" {
+ logger.SetLevel(logrus.DebugLevel)
+ }
+ logger.Formatter = &logrus.JSONFormatter{
+ TimestampFormat: "2006-01-02T15:04:05.000000000Z07:00",
+ }
+ disp := &Dispatcher{logger: logger}
+ err := disp.Run(os.Args[0], os.Args[1:])
if err != nil {
- log.Fatal(err)
+ logrus.Fatalf("%s", err)
}
}
-var (
- theConfig Config
- sqCheck SqueueChecker
-)
-
-const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
+func (disp *Dispatcher) Run(prog string, args []string) error {
+ if err := disp.configure(prog, args); err != nil {
+ return err
+ }
+ disp.setup()
+ return disp.run()
+}
-func doMain() error {
- flags := flag.NewFlagSet("crunch-dispatch-slurm", flag.ExitOnError)
+// configure() loads config files. Tests skip this.
+func (disp *Dispatcher) configure(prog string, args []string) error {
+ flags := flag.NewFlagSet(prog, flag.ExitOnError)
flags.Usage = func() { usage(flags) }
configPath := flags.String(
"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
+ }
- err := readConfig(&theConfig, *configPath)
+ disp.logger.Printf("crunch-dispatch-slurm %s started", version)
+
+ err := disp.readConfig(*configPath)
if err != nil {
return err
}
- if theConfig.CrunchRunCommand == nil {
- theConfig.CrunchRunCommand = []string{"crunch-run"}
+ if disp.CrunchRunCommand == nil {
+ disp.CrunchRunCommand = []string{"crunch-run"}
}
- if theConfig.PollPeriod == 0 {
- theConfig.PollPeriod = arvados.Duration(10 * time.Second)
+ if disp.PollPeriod == 0 {
+ disp.PollPeriod = arvados.Duration(10 * time.Second)
}
- if theConfig.Client.APIHost != "" || theConfig.Client.AuthToken != "" {
+ if disp.Client.APIHost != "" || disp.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", disp.Client.APIHost)
+ os.Setenv("ARVADOS_API_TOKEN", disp.Client.AuthToken)
os.Setenv("ARVADOS_API_HOST_INSECURE", "")
- if theConfig.Client.Insecure {
+ if disp.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(disp.Client.KeepServiceURIs, " "))
os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
} else {
- log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).")
+ disp.logger.Warnf("Client credentials missing from config, so falling back on environment variables (deprecated).")
}
if *dumpConfig {
- log.Fatal(config.DumpAndExit(theConfig))
+ return config.DumpAndExit(disp)
+ }
+
+ siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile)
+ if os.IsNotExist(err) {
+ disp.logger.Warnf("no cluster config (%s), proceeding with no node types defined", err)
+ } else if err != nil {
+ return fmt.Errorf("error loading config: %s", err)
+ } else if disp.cluster, err = siteConfig.GetCluster(""); err != nil {
+ return fmt.Errorf("config error: %s", err)
}
+ return nil
+}
+
+// setup() initializes private fields after configure().
+func (disp *Dispatcher) setup() {
+ if disp.logger == nil {
+ disp.logger = logrus.StandardLogger()
+ }
arv, err := arvadosclient.MakeArvadosClient()
if err != nil {
- log.Printf("Error making Arvados client: %v", err)
- return err
+ disp.logger.Fatalf("Error making Arvados client: %v", err)
}
arv.Retries = 25
- sqCheck = SqueueChecker{Period: time.Duration(theConfig.PollPeriod)}
- defer sqCheck.Stop()
-
- dispatcher := dispatch.Dispatcher{
+ disp.slurm = NewSlurmCLI()
+ disp.sqCheck = &SqueueChecker{
+ Logger: disp.logger,
+ Period: time.Duration(disp.PollPeriod),
+ PrioritySpread: disp.PrioritySpread,
+ Slurm: disp.slurm,
+ }
+ disp.Dispatcher = &dispatch.Dispatcher{
Arv: arv,
- RunContainer: run,
- PollPeriod: time.Duration(theConfig.PollPeriod),
- MinRetryPeriod: time.Duration(theConfig.MinRetryPeriod),
+ Logger: disp.logger,
+ BatchSize: disp.BatchSize,
+ RunContainer: disp.runContainer,
+ PollPeriod: time.Duration(disp.PollPeriod),
+ MinRetryPeriod: time.Duration(disp.MinRetryPeriod),
+ }
+}
+
+func (disp *Dispatcher) run() error {
+ defer disp.sqCheck.Stop()
+
+ if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 {
+ go SlurmNodeTypeFeatureKludge(disp.cluster)
}
if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
log.Printf("Error notifying init daemon: %v", err)
}
-
- return dispatcher.Run()
+ go disp.checkSqueueForOrphans()
+ return disp.Dispatcher.Run(context.Background())
}
-// sbatchCmd
-func sbatchFunc(container arvados.Container) *exec.Cmd {
- memPerCPU := math.Ceil(float64(container.RuntimeConstraints.RAM) / (float64(container.RuntimeConstraints.VCPUs) * 1048576))
-
- var sbatchArgs []string
- sbatchArgs = append(sbatchArgs, "--share")
- 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, ",")))
+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 (disp *Dispatcher) checkSqueueForOrphans() {
+ for _, uuid := range disp.sqCheck.All() {
+ if !containerUuidPattern.MatchString(uuid) {
+ continue
+ }
+ err := disp.TrackContainer(uuid)
+ if err != nil {
+ log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err)
+ }
}
-
- return exec.Command("sbatch", sbatchArgs...)
}
-// scancelCmd
-func scancelFunc(container arvados.Container) *exec.Cmd {
- return exec.Command("scancel", "--name="+container.UUID)
-}
+func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []string {
+ mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM+disp.ReserveExtraRAM) / float64(1048576)))
-// 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
- }
- dispatcher.Unlock(container.UUID)
- }()
+ disk := dispatchcloud.EstimateScratchSpace(&container)
+ disk = int64(math.Ceil(float64(disk) / float64(1048576)))
+ return []string{
+ fmt.Sprintf("--mem=%d", mem),
+ fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs),
+ fmt.Sprintf("--tmp=%d", disk),
+ }
+}
- cmd := sbatchCmd(container)
+func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) {
+ var args []string
+ args = append(args, disp.SbatchArguments...)
+ args = append(args, "--job-name="+container.UUID, fmt.Sprintf("--nice=%d", initialNiceValue), "--no-requeue")
+
+ if disp.cluster == nil {
+ // no instance types configured
+ args = append(args, disp.slurmConstraintArgs(container)...)
+ } else if it, err := dispatchcloud.ChooseInstanceType(disp.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured {
+ // ditto
+ args = append(args, disp.slurmConstraintArgs(container)...)
+ } else if err != nil {
+ return nil, err
+ } else {
+ // use instancetype constraint instead of slurm mem/cpu/tmp specs
+ args = append(args, "--constraint=instancetype="+it.Name)
+ }
- // 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)))
+ if len(container.SchedulingParameters.Partitions) > 0 {
+ args = append(args, "--partition="+strings.Join(container.SchedulingParameters.Partitions, ","))
+ }
- var stdout, stderr bytes.Buffer
- cmd.Stdout = &stdout
- cmd.Stderr = &stderr
+ return args, nil
+}
- // Mutex between squeue sync and running sbatch or scancel.
- sqCheck.L.Lock()
- defer sqCheck.L.Unlock()
+func (disp *Dispatcher) 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))
- 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.Bytes())
- default:
- return fmt.Errorf("exec failed: %v", err)
+ sbArgs, err := disp.sbatchArgs(container)
+ if err != nil {
+ return err
}
+ log.Printf("running sbatch %+q", sbArgs)
+ return disp.slurm.Batch(crScript, sbArgs)
}
-// 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 sqCheck.HasUUID(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)
+// 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 (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
+ ctx, cancel := context.WithCancel(context.Background())
+ defer cancel()
+
+ if ctr.State == dispatch.Locked && !disp.sqCheck.HasUUID(ctr.UUID) {
+ log.Printf("Submitting container %s to slurm", ctr.UUID)
+ if err := disp.submit(ctr, disp.CrunchRunCommand); err != nil {
+ var text string
+ if err, ok := err.(dispatchcloud.ConstraintsNotSatisfiableError); ok {
+ var logBuf bytes.Buffer
+ fmt.Fprintf(&logBuf, "cannot run container %s: %s\n", ctr.UUID, err)
+ if len(err.AvailableTypes) == 0 {
+ fmt.Fprint(&logBuf, "No instance types are configured.\n")
+ } else {
+ fmt.Fprint(&logBuf, "Available instance types:\n")
+ for _, t := range err.AvailableTypes {
+ fmt.Fprintf(&logBuf,
+ "Type %q: %d VCPUs, %d RAM, %d Scratch, %f Price\n",
+ t.Name, t.VCPUs, t.RAM, t.Scratch, t.Price,
+ )
+ }
+ }
+ text = logBuf.String()
+ disp.UpdateState(ctr.UUID, dispatch.Cancelled)
+ } else {
+ text = fmt.Sprintf("Error submitting container %s to slurm: %s", ctr.UUID, err)
}
+ log.Print(text)
- 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)
- 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
- }
+ lr := arvadosclient.Dict{"log": arvadosclient.Dict{
+ "object_uuid": ctr.UUID,
+ "event_type": "dispatch",
+ "properties": map[string]string{"text": text}}}
+ disp.Arv.Create("logs", lr, nil)
+
+ disp.Unlock(ctr.UUID)
+ return
}
}
-}
-// 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)
-
- monitorDone := false
- go monitorSubmitOrCancel(dispatcher, container, &monitorDone)
-
- for container = range status {
- 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()
-
- 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
+ log.Printf("Start monitoring container %v in state %q", ctr.UUID, ctr.State)
+ defer log.Printf("Done monitoring container %s", ctr.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 && disp.sqCheck.HasUUID(uuid) {
+ }
+ cancel()
+ }(ctr.UUID)
+
+ for {
+ select {
+ case <-ctx.Done():
+ // Disappeared from squeue
+ if err := disp.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:
+ disp.UpdateState(ctr.UUID, dispatch.Cancelled)
+ case dispatch.Locked:
+ disp.Unlock(ctr.UUID)
+ }
+ return
+ case updated, ok := <-status:
+ if !ok {
+ log.Printf("container %s is done: cancel slurm job", ctr.UUID)
+ disp.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)
+ disp.scancel(ctr)
+ } else {
+ p := int64(updated.Priority)
+ if p <= 1000 {
+ // API is providing
+ // user-assigned priority. If
+ // ctrs have equal priority,
+ // run the older one first.
+ p = int64(p)<<50 - (updated.CreatedAt.UnixNano() >> 14)
}
+ disp.sqCheck.SetPriority(ctr.UUID, p)
}
-
- // Ignore errors; if necessary, we'll try again next time
- dispatcher.UpdateState(container.UUID, dispatch.Cancelled)
}
}
- monitorDone = true
+}
+func (disp *Dispatcher) scancel(ctr arvados.Container) {
+ err := disp.slurm.Cancel(ctr.UUID)
+ if err != nil {
+ log.Printf("scancel: %s", err)
+ time.Sleep(time.Second)
+ } else if disp.sqCheck.HasUUID(ctr.UUID) {
+ log.Printf("container %s is still in squeue after scancel", ctr.UUID)
+ time.Sleep(time.Second)
+ }
}
-func readConfig(dst interface{}, path string) error {
- err := config.LoadFile(dst, path)
+func (disp *Dispatcher) readConfig(path string) error {
+ err := config.LoadFile(disp, path)
if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
log.Printf("Config not specified. Continue with default configuration.")
err = nil