X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/10f08c358c12468119dc2621c48b68d6d33417da..7db74f672f64b3e647a98c1d8e5978b50d79538d:/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 879cb785d9..1c0f6ad28f 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -2,163 +2,158 @@ // // SPDX-License-Identifier: AGPL-3.0 -package main - // Dispatcher service for Crunch that submits containers to the slurm queue. +package dispatchslurm import ( "context" - "flag" + "crypto/hmac" + "crypto/sha256" "fmt" "log" "math" + "net/http" "os" "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" + "git.arvados.org/arvados.git/lib/cmd" + "git.arvados.org/arvados.git/lib/controller/dblock" + "git.arvados.org/arvados.git/lib/ctrlctx" + "git.arvados.org/arvados.git/lib/dispatchcloud" + "git.arvados.org/arvados.git/lib/service" + "git.arvados.org/arvados.git/sdk/go/arvados" + "git.arvados.org/arvados.git/sdk/go/arvadosclient" + "git.arvados.org/arvados.git/sdk/go/ctxlog" + "git.arvados.org/arvados.git/sdk/go/dispatch" "github.com/coreos/go-systemd/daemon" + "github.com/prometheus/client_golang/prometheus" + "github.com/sirupsen/logrus" ) -var version = "dev" +var Command cmd.Handler = service.Command(arvados.ServiceNameDispatchSLURM, newHandler) -type command struct { - dispatcher *dispatch.Dispatcher - cluster *arvados.Cluster - sqCheck *SqueueChecker - slurm Slurm +func newHandler(ctx context.Context, cluster *arvados.Cluster, _ string, _ *prometheus.Registry) service.Handler { + logger := ctxlog.FromContext(ctx) + disp := &Dispatcher{logger: logger, cluster: cluster} + if err := disp.configure(); err != nil { + return service.ErrorHandler(ctx, cluster, err) + } + disp.setup() + go func() { + disp.err = disp.run() + close(disp.done) + }() + return disp +} - Client arvados.Client +type logger interface { + dispatch.Logger + Fatalf(string, ...interface{}) +} + +const initialNiceValue int64 = 10000 - SbatchArguments []string - PollPeriod arvados.Duration +type Dispatcher struct { + *dispatch.Dispatcher + logger logrus.FieldLogger + cluster *arvados.Cluster + sqCheck *SqueueChecker + slurm Slurm + dbConnector ctrlctx.DBConnector - // 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 + done chan struct{} + err error - // Minimum time between two attempts to run the same container - MinRetryPeriod arvados.Duration + Client arvados.Client } -func main() { - err := (&command{}).Run(os.Args[0], os.Args[1:]) - if err != nil { - log.Fatal(err) - } +func (disp *Dispatcher) CheckHealth() error { + return disp.err } -const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml" - -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(args) - - // Print version information if requested - if *getVersion { - fmt.Printf("crunch-dispatch-slurm %s\n", version) - return nil - } +func (disp *Dispatcher) Done() <-chan struct{} { + return disp.done +} - log.Printf("crunch-dispatch-slurm %s started", version) +func (disp *Dispatcher) ServeHTTP(w http.ResponseWriter, r *http.Request) { + http.NotFound(w, r) +} - err := cmd.readConfig(*configPath) - if err != nil { - return err +// configure() loads config files. Some tests skip this (see +// StubbedSuite). +func (disp *Dispatcher) configure() error { + if disp.logger == nil { + disp.logger = logrus.StandardLogger() } + disp.logger = disp.logger.WithField("ClusterID", disp.cluster.ClusterID) + disp.logger.Printf("crunch-dispatch-slurm %s started", cmd.Version.String()) - if cmd.CrunchRunCommand == nil { - cmd.CrunchRunCommand = []string{"crunch-run"} - } + disp.Client.APIHost = disp.cluster.Services.Controller.ExternalURL.Host + disp.Client.AuthToken = disp.cluster.SystemRootToken + disp.Client.Insecure = disp.cluster.TLS.Insecure + disp.dbConnector = ctrlctx.DBConnector{PostgreSQL: disp.cluster.PostgreSQL} - if cmd.PollPeriod == 0 { - cmd.PollPeriod = arvados.Duration(10 * time.Second) - } - - if cmd.Client.APIHost != "" || cmd.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", cmd.Client.APIHost) - os.Setenv("ARVADOS_API_TOKEN", cmd.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 cmd.Client.Insecure { + if disp.Client.Insecure { os.Setenv("ARVADOS_API_HOST_INSECURE", "1") } - os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(cmd.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).") - } - - if *dumpConfig { - log.Fatal(config.DumpAndExit(cmd)) + disp.logger.Warnf("Client credentials missing from config, so falling back on environment variables (deprecated).") } + return nil +} +// setup() initializes private fields after configure(). +func (disp *Dispatcher) setup() { + disp.done = make(chan struct{}) 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 - 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) + 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, } - - if cmd.slurm == nil { - cmd.slurm = &slurmCLI{} + disp.Dispatcher = &dispatch.Dispatcher{ + Arv: arv, + 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), } +} - cmd.sqCheck = &SqueueChecker{ - Period: time.Duration(cmd.PollPeriod), - Slurm: cmd.slurm, - } - defer cmd.sqCheck.Stop() +func (disp *Dispatcher) run() error { + dblock.Dispatch.Lock(context.Background(), disp.dbConnector.GetDB) + defer dblock.Dispatch.Unlock() + defer disp.sqCheck.Stop() - cmd.dispatcher = &dispatch.Dispatcher{ - Arv: arv, - RunContainer: cmd.run, - PollPeriod: time.Duration(cmd.PollPeriod), - MinRetryPeriod: time.Duration(cmd.MinRetryPeriod), + 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) } - - go cmd.checkSqueueForOrphans() - - return cmd.dispatcher.Run(context.Background()) + go disp.checkSqueueForOrphans() + return disp.Dispatcher.Run(context.Background()) } var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$`) @@ -168,110 +163,93 @@ var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$` // 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) { +func (disp *Dispatcher) checkSqueueForOrphans() { + for _, uuid := range disp.sqCheck.All() { + if !containerUuidPattern.MatchString(uuid) || !strings.HasPrefix(uuid, disp.cluster.ClusterID) { continue } - err := cmd.dispatcher.TrackContainer(uuid) + err := disp.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 -} - -func (cmd *command) sbatchArgs(container arvados.Container) ([]string, error) { - mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM) / float64(1048576))) +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))) - 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))) - - var sbatchArgs []string - sbatchArgs = append(sbatchArgs, cmd.SbatchArguments...) - sbatchArgs = append(sbatchArgs, fmt.Sprintf("--job-name=%s", container.UUID)) - sbatchArgs = append(sbatchArgs, fmt.Sprintf("--mem=%d", mem)) - sbatchArgs = append(sbatchArgs, fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs)) - 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 []string{ + fmt.Sprintf("--mem=%d", mem), + fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs), + fmt.Sprintf("--tmp=%d", disk), } +} - if cmd.cluster == nil { +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 - } else if it, err := dispatchcloud.ChooseInstanceType(cmd.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured { + 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 { - sbatchArgs = append(sbatchArgs, "--constraint="+it.Name) + // use instancetype constraint instead of slurm mem/cpu/tmp specs + args = append(args, "--constraint=instancetype="+it.Name) + } + + if len(container.SchedulingParameters.Partitions) > 0 { + args = append(args, "--partition="+strings.Join(container.SchedulingParameters.Partitions, ",")) } - return sbatchArgs, nil + return args, nil } -func (cmd *command) submit(container arvados.Container, crunchRunCommand []string) error { +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, "--runtime-engine="+disp.cluster.Containers.RuntimeEngine) crArgs = append(crArgs, container.UUID) - crScript := strings.NewReader(execScript(crArgs)) - cmd.sqCheck.L.Lock() - defer cmd.sqCheck.L.Unlock() + h := hmac.New(sha256.New, []byte(disp.cluster.SystemRootToken)) + fmt.Fprint(h, container.UUID) + authsecret := fmt.Sprintf("%x", h.Sum(nil)) - sbArgs, err := cmd.sbatchArgs(container) + crScript := strings.NewReader(execScript(crArgs, map[string]string{"GatewayAuthSecret": authsecret})) + + sbArgs, err := disp.sbatchArgs(container) if err != nil { return err } log.Printf("running sbatch %+q", sbArgs) - return cmd.slurm.Batch(crScript, 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 (cmd *command) run(_ *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 && !cmd.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 := 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) - - 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) - - cmd.dispatcher.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 } } @@ -282,7 +260,7 @@ func (cmd *command) run(_ *dispatch.Dispatcher, ctr arvados.Container, status <- // 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) { + for ctx.Err() == nil && disp.sqCheck.HasUUID(uuid) { } cancel() }(ctr.UUID) @@ -291,71 +269,44 @@ func (cmd *command) run(_ *dispatch.Dispatcher, ctr arvados.Container, status <- 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) + 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: - cmd.dispatcher.UpdateState(ctr.UUID, dispatch.Cancelled) + disp.UpdateState(ctr.UUID, dispatch.Cancelled) case dispatch.Locked: - cmd.dispatcher.Unlock(ctr.UUID) + disp.Unlock(ctr.UUID) } - return + return nil case updated, ok := <-status: if !ok { - log.Printf("Dispatcher says container %s is done: cancel slurm job", ctr.UUID) - cmd.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) - cmd.scancel(ctr) + log.Printf("container %s has state %q, priority %d: cancel slurm job", ctr.UUID, updated.State, updated.Priority) + disp.scancel(ctr) } else { - cmd.renice(updated) + 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 (cmd *command) scancel(ctr arvados.Container) { - cmd.sqCheck.L.Lock() - err := cmd.slurm.Cancel(ctr.UUID) - cmd.sqCheck.L.Unlock() - +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 cmd.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 (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() - - 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)) - } -} - -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 - } - return err -}