X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/9f9cf2e1da7233d6b341602faeaa69454b7e201f..37d9f94b06ff367a3514b58ec6f0e4d4d0116030:/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 b4103cc625..ff1077fae6 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -7,7 +7,6 @@ package main // Dispatcher service for Crunch that submits containers to the slurm queue. import ( - "bytes" "context" "flag" "fmt" @@ -18,52 +17,50 @@ import ( "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" + "git.arvados.org/arvados.git/lib/cmd" + "git.arvados.org/arvados.git/lib/config" + "git.arvados.org/arvados.git/lib/dispatchcloud" + "git.arvados.org/arvados.git/sdk/go/arvados" + "git.arvados.org/arvados.git/sdk/go/arvadosclient" + "git.arvados.org/arvados.git/sdk/go/dispatch" "github.com/coreos/go-systemd/daemon" + "github.com/ghodss/yaml" + "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" + version = "dev" ) 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 } func main() { - disp := &Dispatcher{} + 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) } } @@ -77,13 +74,15 @@ func (disp *Dispatcher) Run(prog string, args []string) error { // configure() loads config files. Tests skip this. func (disp *Dispatcher) configure(prog string, args []string) error { - flags := flag.NewFlagSet(prog, flag.ExitOnError) + if disp.logger == nil { + disp.logger = logrus.StandardLogger() + } + flags := flag.NewFlagSet(prog, flag.ContinueOnError) 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, @@ -92,8 +91,11 @@ func (disp *Dispatcher) configure(prog string, args []string) error { "version", false, "Print version information and exit.") - // Parse args; omit the first arg which is the command name - flags.Parse(args) + + args = loader.MungeLegacyConfigArgs(disp.logger, args, "-legacy-crunch-dispatch-slurm-config") + if ok, code := cmd.ParseFlags(flags, prog, args, "", os.Stderr); !ok { + os.Exit(code) + } // Print version information if requested if *getVersion { @@ -101,20 +103,20 @@ func (disp *Dispatcher) configure(prog string, args []string) error { return nil } - log.Printf("crunch-dispatch-slurm %s started", version) + disp.logger.Printf("crunch-dispatch-slurm %s started", version) - err := disp.readConfig(*configPath) + cfg, err := loader.Load() if err != nil { return err } - if disp.CrunchRunCommand == nil { - disp.CrunchRunCommand = []string{"crunch-run"} + if disp.cluster, err = cfg.GetCluster(""); err != nil { + return fmt.Errorf("config error: %s", err) } - if disp.PollPeriod == 0 { - disp.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 disp.Client.APIHost != "" || disp.Client.AuthToken != "" { // Copy real configs into env vars so [a] @@ -126,23 +128,23 @@ func (disp *Dispatcher) configure(prog string, args []string) error { if disp.Client.Insecure { os.Setenv("ARVADOS_API_HOST_INSECURE", "1") } - os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(disp.Client.KeepServiceURIs, " ")) os.Setenv("ARVADOS_EXTERNAL_CLIENT", "") + for k, v := range disp.cluster.Containers.SLURM.SbatchEnvironmentVariables { + os.Setenv(k, v) + } } 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 { - return config.DumpAndExit(disp) - } - - siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile) - if os.IsNotExist(err) { - log.Printf("warning: 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) + out, err := yaml.Marshal(cfg) + if err != nil { + return err + } + _, err = os.Stdout.Write(out) + if err != nil { + return err + } } return nil @@ -152,21 +154,24 @@ func (disp *Dispatcher) configure(prog string, args []string) error { func (disp *Dispatcher) setup() { arv, err := arvadosclient.MakeArvadosClient() if err != nil { - log.Fatalf("Error making Arvados client: %v", err) + disp.logger.Fatalf("Error making Arvados client: %v", err) } arv.Retries = 25 - disp.slurm = &slurmCLI{} + disp.slurm = NewSlurmCLI() disp.sqCheck = &SqueueChecker{ - Period: time.Duration(disp.PollPeriod), - PrioritySpread: disp.PrioritySpread, + 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, + Logger: disp.logger, + BatchSize: disp.cluster.API.MaxItemsPerResponse, RunContainer: disp.runContainer, - PollPeriod: time.Duration(disp.PollPeriod), - MinRetryPeriod: time.Duration(disp.MinRetryPeriod), + PollPeriod: time.Duration(disp.cluster.Containers.CloudVMs.PollInterval), + MinRetryPeriod: time.Duration(disp.cluster.Containers.MinRetryPeriod), } } @@ -174,7 +179,7 @@ func (disp *Dispatcher) run() error { defer disp.sqCheck.Stop() if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 { - go dispatchcloud.SlurmNodeTypeFeatureKludge(disp.cluster) + go SlurmNodeTypeFeatureKludge(disp.cluster) } if _, err := daemon.SdNotify(false, "READY=1"); err != nil { @@ -193,7 +198,7 @@ var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$` // Cancelled or Complete. See https://dev.arvados.org/issues/10979 func (disp *Dispatcher) checkSqueueForOrphans() { for _, uuid := range disp.sqCheck.All() { - if !containerUuidPattern.MatchString(uuid) { + if !containerUuidPattern.MatchString(uuid) || !strings.HasPrefix(uuid, disp.cluster.ClusterID) { continue } err := disp.TrackContainer(uuid) @@ -204,14 +209,11 @@ func (disp *Dispatcher) checkSqueueForOrphans() { } func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []string { - mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM+disp.ReserveExtraRAM) / float64(1048576))) + mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+ + container.RuntimeConstraints.KeepCacheRAM+ + int64(disp.cluster.Containers.ReserveExtraRAM)) / float64(1048576))) - var disk int64 - for _, m := range container.Mounts { - if m.Kind == "tmp" { - disk += m.Capacity - } - } + disk := dispatchcloud.EstimateScratchSpace(&container) disk = int64(math.Ceil(float64(disk) / float64(1048576))) return []string{ fmt.Sprintf("--mem=%d", mem), @@ -222,8 +224,8 @@ func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []strin 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)) + 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 @@ -249,6 +251,7 @@ func (disp *Dispatcher) submit(container arvados.Container, crunchRunCommand []s // append() here avoids modifying crunchRunCommand's // underlying array, which is shared with other goroutines. crArgs := append([]string(nil), crunchRunCommand...) + crArgs = append(crArgs, "--runtime-engine="+disp.cluster.Containers.RuntimeEngine) crArgs = append(crArgs, container.UUID) crScript := strings.NewReader(execScript(crArgs)) @@ -264,43 +267,17 @@ func (disp *Dispatcher) submit(container arvados.Container, crunchRunCommand []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) { +func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) error { 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) - - 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 + cmd := []string{disp.cluster.Containers.CrunchRunCommand} + cmd = append(cmd, disp.cluster.Containers.CrunchRunArgumentsList...) + err := disp.submit(ctr, cmd) + if err != nil { + return err } } @@ -321,7 +298,7 @@ func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Contain 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: @@ -329,7 +306,7 @@ func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Contain case dispatch.Locked: disp.Unlock(ctr.UUID) } - return + return nil case updated, ok := <-status: if !ok { log.Printf("container %s is done: cancel slurm job", ctr.UUID) @@ -361,12 +338,3 @@ func (disp *Dispatcher) scancel(ctr arvados.Container) { time.Sleep(time.Second) } } - -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 - } - return err -}