X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/dc7d01f4d4031962ffd5734ca0c64146a7217e4a..707ed2964a24ebe940ae440af22fdcb56781fcbd:/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 f1f5d1e6b8..0879de20f9 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -1,19 +1,23 @@ +// 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 ( + "context" "flag" "fmt" - "io" - "io/ioutil" "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" @@ -21,8 +25,17 @@ import ( "github.com/coreos/go-systemd/daemon" ) -// Config used by crunch-dispatch-slurm -type Config struct { +var ( + version = "dev" + defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml" +) + +type Dispatcher struct { + *dispatch.Dispatcher + cluster *arvados.Cluster + sqCheck *SqueueChecker + slurm Slurm + Client arvados.Client SbatchArguments []string @@ -33,24 +46,30 @@ type Config struct { // // Example: []string{"crunch-run", "--cgroup-parent-subsystem=memory"} CrunchRunCommand []string + + // Minimum time between two attempts to run the same container + MinRetryPeriod arvados.Duration } func main() { - err := doMain() + disp := &Dispatcher{} + err := disp.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 (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( @@ -61,278 +80,293 @@ func doMain() error { "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 := 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).") } if *dumpConfig { - log.Fatal(config.DumpAndExit(theConfig)) + 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) + } + + 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 + log.Fatalf("Error making Arvados client: %v", err) } arv.Retries = 25 - squeueUpdater.StartMonitor(time.Duration(theConfig.PollPeriod)) - defer squeueUpdater.Done() - - dispatcher := dispatch.Dispatcher{ - Arv: arv, - RunContainer: run, - PollInterval: time.Duration(theConfig.PollPeriod), - DoneProcessing: make(chan struct{})} - - if _, err := daemon.SdNotify(false, "READY=1"); err != nil { - log.Printf("Error notifying init daemon: %v", err) + disp.slurm = &slurmCLI{} + disp.sqCheck = &SqueueChecker{ + Period: time.Duration(disp.PollPeriod), + Slurm: disp.slurm, } - - err = dispatcher.RunDispatcher() - if err != nil { - return err + disp.Dispatcher = &dispatch.Dispatcher{ + Arv: arv, + RunContainer: disp.runContainer, + PollPeriod: time.Duration(disp.PollPeriod), + MinRetryPeriod: time.Duration(disp.MinRetryPeriod), } - - return nil } -// sbatchCmd -func sbatchFunc(container arvados.Container) *exec.Cmd { - memPerCPU := math.Ceil(float64(container.RuntimeConstraints.RAM) / (float64(container.RuntimeConstraints.VCPUs) * 1048576)) +func (disp *Dispatcher) run() error { + defer disp.sqCheck.Stop() - 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, ","))) + if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 { + go dispatchcloud.SlurmNodeTypeFeatureKludge(disp.cluster) } - return exec.Command("sbatch", sbatchArgs...) -} - -// scancelCmd -func scancelFunc(container arvados.Container) *exec.Cmd { - return exec.Command("scancel", "--name="+container.UUID) + if _, err := daemon.SdNotify(false, "READY=1"); err != nil { + log.Printf("Error notifying init daemon: %v", err) + } + go disp.checkSqueueForOrphans() + return disp.Dispatcher.Run(context.Background()) } -// 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 +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 := dispatcher.Unlock(container.UUID) + err := disp.TrackContainer(uuid) if err != nil { - log.Printf("Error unlocking container %s: %v", container.UUID, err) + log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err) } - }() - - // Create the command and attach to stdin/stdout - cmd := sbatchCmd(container) - stdinWriter, stdinerr := cmd.StdinPipe() - if stdinerr != nil { - submitErr = fmt.Errorf("Error creating stdin pipe %v: %q", container.UUID, stdinerr) - return } +} - stdoutReader, stdoutErr := cmd.StdoutPipe() - if stdoutErr != nil { - submitErr = fmt.Errorf("Error creating stdout pipe %v: %q", container.UUID, stdoutErr) - return +func (disp *Dispatcher) niceness(priority int) int { + if priority > 1000 { + priority = 1000 } - - stderrReader, stderrErr := cmd.StderrPipe() - if stderrErr != nil { - submitErr = fmt.Errorf("Error creating stderr pipe %v: %q", container.UUID, stderrErr) - return + if priority < 0 { + priority = 0 } + // Niceness range 1-10000 + return (1000 - priority) * 10 +} - // Mutex between squeue sync and running sbatch or scancel. - squeueUpdater.SlurmLock.Lock() - defer squeueUpdater.SlurmLock.Unlock() +func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) { + mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM) / float64(1048576))) - log.Printf("sbatch starting: %+q", cmd.Args) - err := cmd.Start() - if err != nil { - submitErr = fmt.Errorf("Error starting sbatch: %v", err) - return + var disk int64 + for _, m := range container.Mounts { + if m.Kind == "tmp" { + disk += m.Capacity + } } + disk = int64(math.Ceil(float64(disk) / float64(1048576))) - stdoutChan := make(chan []byte) - go func() { - b, _ := ioutil.ReadAll(stdoutReader) - stdoutReader.Close() - stdoutChan <- b - close(stdoutChan) - }() + var sbatchArgs []string + sbatchArgs = append(sbatchArgs, disp.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", disp.niceness(container.Priority))) + if len(container.SchedulingParameters.Partitions) > 0 { + sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ","))) + } - stderrChan := make(chan []byte) - go func() { - b, _ := ioutil.ReadAll(stderrReader) - stderrReader.Close() - stderrChan <- b - close(stderrChan) - }() + if disp.cluster == nil { + // no instance types configured + } else if it, err := dispatchcloud.ChooseInstanceType(disp.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured { + // ditto + } else if err != nil { + return nil, err + } else { + sbatchArgs = append(sbatchArgs, "--constraint=instancetype="+it.Name) + } - // Send a tiny script on stdin to execute the crunch-run command - // slurm actually enforces that this must be a #! script - io.WriteString(stdinWriter, execScript(append(crunchRunCommand, container.UUID))) - stdinWriter.Close() + return sbatchArgs, nil +} - stdoutMsg := <-stdoutChan - stderrmsg := <-stderrChan +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)) - err = cmd.Wait() + disp.sqCheck.L.Lock() + defer disp.sqCheck.L.Unlock() + sbArgs, err := disp.sbatchArgs(container) if err != nil { - submitErr = fmt.Errorf("Container submission failed: %v: %v (stderr: %q)", cmd.Args, err, stderrmsg) - return + return err } - - log.Printf("sbatch succeeded: %s", strings.TrimSpace(string(stdoutMsg))) - return + 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 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) +// 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 == dispatchcloud.ErrConstraintsNotSatisfiable { + text = fmt.Sprintf("cannot run container %s: %s", ctr.UUID, err) + 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) + 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 %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: - 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 + 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 { + disp.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) - - monitorDone := false - go monitorSubmitOrCancel(dispatcher, container, &monitorDone) - - for container = range status { - if container.State == dispatch.Locked || container.State == dispatch.Running { - if container.Priority == 0 { - 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("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 - } - } - - err = dispatcher.UpdateState(container.UUID, dispatch.Cancelled) - } - } +func (disp *Dispatcher) scancel(ctr arvados.Container) { + disp.sqCheck.L.Lock() + err := disp.slurm.Cancel(ctr.UUID) + disp.sqCheck.L.Unlock() + + 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 (disp *Dispatcher) renice(ctr arvados.Container) { + nice := disp.niceness(ctr.Priority) + oldnice := disp.sqCheck.GetNiceness(ctr.UUID) + if nice == oldnice || oldnice == -1 { + return + } + log.Printf("updating slurm nice value to %d (was %d)", nice, oldnice) + disp.sqCheck.L.Lock() + err := disp.slurm.Renice(ctr.UUID, nice) + disp.sqCheck.L.Unlock() + + if err != nil { + log.Printf("renice: %s", err) + time.Sleep(time.Second) + return + } + if disp.sqCheck.HasUUID(ctr.UUID) { + log.Printf("container %s has arvados priority %d, slurm nice %d", + ctr.UUID, ctr.Priority, disp.sqCheck.GetNiceness(ctr.UUID)) } - monitorDone = true } -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