X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/4cd89bd1767bece226c412ae7c9ea37669e8706b..dbbb649c9b797da8f6805af3a7e74e57b075d294:/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 617b076da2..300e923640 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -1,3 +1,7 @@ +// 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. @@ -10,206 +14,319 @@ import ( "log" "math" "os" - "os/exec" + "regexp" "strings" "time" + "git.curoverse.com/arvados.git/lib/config" + "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/ghodss/yaml" + "github.com/sirupsen/logrus" ) -// Config used by crunch-dispatch-slurm -type Config struct { - Client arvados.Client +type logger interface { + dispatch.Logger + Fatalf(string, ...interface{}) +} - SbatchArguments []string - PollPeriod arvados.Duration +const initialNiceValue int64 = 10000 - // 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 +var ( + version = "dev" +) - // Minimum time between two attempts to run the same container - MinRetryPeriod arvados.Duration +type Dispatcher struct { + *dispatch.Dispatcher + logger logrus.FieldLogger + cluster *arvados.Cluster + sqCheck *SqueueChecker + slurm Slurm + + Client arvados.Client } 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 { + if disp.logger == nil { + disp.logger = logrus.StandardLogger() + } + flags := flag.NewFlagSet(prog, flag.ExitOnError) flags.Usage = func() { usage(flags) } - configPath := flags.String( - "config", - defaultConfigPath, - "`path` to JSON or YAML configuration file") + loader := config.NewLoader(nil, disp.logger) + loader.SetupFlags(flags) + dumpConfig := flag.Bool( "dump-config", false, "write current configuration to stdout and exit") + getVersion := flags.Bool( + "version", + false, + "Print version information and exit.") + + args = loader.MungeLegacyConfigArgs(logrus.StandardLogger(), args, "-legacy-crunch-dispatch-slurm-config") // Parse args; omit the first arg which is the command name - flags.Parse(os.Args[1:]) + err := flags.Parse(args) + + if err == flag.ErrHelp { + return nil + } + + // Print version information if requested + if *getVersion { + fmt.Printf("crunch-dispatch-slurm %s\n", version) + return nil + } + + disp.logger.Printf("crunch-dispatch-slurm %s started", version) - err := readConfig(&theConfig, *configPath) + cfg, err := loader.Load() if err != nil { return err } - if theConfig.CrunchRunCommand == nil { - theConfig.CrunchRunCommand = []string{"crunch-run"} + if disp.cluster, err = cfg.GetCluster(""); err != nil { + return fmt.Errorf("config error: %s", err) } - if theConfig.PollPeriod == 0 { - theConfig.PollPeriod = arvados.Duration(10 * time.Second) - } + disp.Client.APIHost = disp.cluster.Services.Controller.ExternalURL.Host + disp.Client.AuthToken = disp.cluster.SystemRootToken + disp.Client.Insecure = disp.cluster.TLS.Insecure - 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, " ")) + ks := "" + if length(disp.cluster.Containers.SLURM.KeepServices) > 0 { + for _, svc := range disp.cluster.Containers.SLURM.KeepServices { + for k, _ := range svc.InternalURLs { + ks += k + ks += " " + } + } + } + os.Setenv("ARVADOS_KEEP_SERVICES", ks) 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)) + out, err := yaml.Marshal(cfg) + if err != nil { + return err + } + _, err = os.Stdout.Write(out) + if err != nil { + return err + } } + return nil +} + +// setup() initializes private fields after configure(). +func (disp *Dispatcher) setup() { 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.cluster.Containers.CloudVMs.PollInterval), + PrioritySpread: disp.cluster.Containers.SLURM.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.cluster.API.MaxItemsPerResponse, + RunContainer: disp.runContainer, + PollPeriod: time.Duration(disp.cluster.Containers.CloudVMs.PollInterval), + MinRetryPeriod: time.Duration(disp.cluster.Containers.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(context.Background()) + 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 len(container.SchedulingParameters.Partitions) > 0 { - 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+ + int64(disp.cluster.Containers.ReserveExtraRAM)) / float64(1048576))) + + 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), + } } -// 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) error { - cmd := sbatchCmd(container) - - // 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))) - - var stdout, stderr bytes.Buffer - cmd.Stdout = &stdout - cmd.Stderr = &stderr - - // Mutex between squeue sync and running sbatch or scancel. - sqCheck.L.Lock() - defer sqCheck.L.Unlock() +func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) { + var args []string + args = append(args, disp.cluster.Containers.SLURM.SbatchArgumentsList...) + 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) + } - log.Printf("exec sbatch %+q", cmd.Args) - err := cmd.Run() + if len(container.SchedulingParameters.Partitions) > 0 { + args = append(args, "--partition="+strings.Join(container.SchedulingParameters.Partitions, ",")) + } - switch err.(type) { - case nil: - log.Printf("sbatch succeeded: %q", strings.TrimSpace(stdout.String())) - return nil + return args, nil +} - case *exec.ExitError: - dispatcher.Unlock(container.UUID) - return fmt.Errorf("sbatch %+q failed: %v (stderr: %q)", cmd.Args, err, stderr.Bytes()) +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)) - default: - dispatcher.Unlock(container.UUID) - 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) } // 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 run(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) { +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 && !sqCheck.HasUUID(ctr.UUID) { + if ctr.State == dispatch.Locked && !disp.sqCheck.HasUUID(ctr.UUID) { log.Printf("Submitting container %s to slurm", ctr.UUID) - if err := submit(disp, ctr, theConfig.CrunchRunCommand); err != nil { - log.Printf("Error submitting container %s to slurm: %s", ctr.UUID, err) + cmd := []string{disp.cluster.Containers.CrunchRunCommand} + cmd = append(cmd, disp.cluster.Containers.CrunchRunArgumentsList...) + if err := disp.submit(ctr, cmd); 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) + + 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 } } - log.Printf("Start monitoring container %s", ctr.UUID) + 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 && sqCheck.HasUUID(uuid) { + for ctx.Err() == nil && disp.sqCheck.HasUUID(uuid) { } cancel() }(ctr.UUID) @@ -219,7 +336,7 @@ func run(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados 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) + log.Printf("error getting final container state for %s: %s", ctr.UUID, err) } switch ctr.State { case dispatch.Running: @@ -230,36 +347,32 @@ func run(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados return case updated, ok := <-status: if !ok { - log.Printf("Dispatcher says container %s is done: cancel slurm job", ctr.UUID) - scancel(ctr) + 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) - scancel(ctr) + 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) } } } } - -func scancel(ctr arvados.Container) { - sqCheck.L.Lock() - cmd := scancelCmd(ctr) - msg, err := cmd.CombinedOutput() - sqCheck.L.Unlock() - +func (disp *Dispatcher) scancel(ctr arvados.Container) { + err := disp.slurm.Cancel(ctr.UUID) if err != nil { - log.Printf("%q %q: %s %q", cmd.Path, cmd.Args, err, msg) + log.Printf("scancel: %s", err) time.Sleep(time.Second) - } else if sqCheck.HasUUID(ctr.UUID) { + } 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) - if err != nil && os.IsNotExist(err) && path == defaultConfigPath { - log.Printf("Config not specified. Continue with default configuration.") - err = nil - } - return err -}