X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/2efdb0e205abe63c5ca777fc4dacca65192e5dd3..dfc4c2c74dc7a31ff0c0d307846b7e9525fdcf0e:/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 f718fbcdce..30cbb79dc1 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -1,49 +1,111 @@ +// 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" - "git.curoverse.com/arvados.git/sdk/go/arvadosclient" - "git.curoverse.com/arvados.git/sdk/go/dispatch" - "io/ioutil" "log" "math" "os" "os/exec" + "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" + "github.com/coreos/go-systemd/daemon" ) +// Config used by crunch-dispatch-slurm +type Config struct { + Client arvados.Client + + SbatchArguments []string + PollPeriod arvados.Duration + + // 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 + + // Minimum time between two attempts to run the same container + MinRetryPeriod arvados.Duration +} + func main() { err := doMain() if err != nil { - log.Fatalf("%q", err) + log.Fatal(err) } } var ( - crunchRunCommand *string - squeueUpdater Squeue + theConfig Config + sqCheck = &SqueueChecker{} ) +const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml" + func doMain() error { flags := flag.NewFlagSet("crunch-dispatch-slurm", flag.ExitOnError) + flags.Usage = func() { usage(flags) } - pollInterval := flags.Int( - "poll-interval", - 10, - "Interval in seconds to poll for queued containers") - - crunchRunCommand = flags.String( - "crunch-run-command", - "/usr/bin/crunch-run", - "Crunch command to run container") + 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") // Parse args; omit the first arg which is the command name flags.Parse(os.Args[1:]) + err := readConfig(&theConfig, *configPath) + if err != nil { + return err + } + + if theConfig.CrunchRunCommand == nil { + theConfig.CrunchRunCommand = []string{"crunch-run"} + } + + if theConfig.PollPeriod == 0 { + theConfig.PollPeriod = arvados.Duration(10 * time.Second) + } + + if theConfig.Client.APIHost != "" || theConfig.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_INSECURE", "") + if theConfig.Client.Insecure { + os.Setenv("ARVADOS_API_HOST_INSECURE", "1") + } + os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(theConfig.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)) + } + arv, err := arvadosclient.MakeArvadosClient() if err != nil { log.Printf("Error making Arvados client: %v", err) @@ -51,35 +113,71 @@ func doMain() error { } arv.Retries = 25 - squeueUpdater.StartMonitor(time.Duration(*pollInterval) * time.Second) - defer squeueUpdater.Done() + sqCheck = &SqueueChecker{Period: time.Duration(theConfig.PollPeriod)} + defer sqCheck.Stop() - dispatcher := dispatch.Dispatcher{ + dispatcher := &dispatch.Dispatcher{ Arv: arv, RunContainer: run, - PollInterval: time.Duration(*pollInterval) * time.Second, - DoneProcessing: make(chan struct{})} + PollPeriod: time.Duration(theConfig.PollPeriod), + MinRetryPeriod: time.Duration(theConfig.MinRetryPeriod), + } - err = dispatcher.RunDispatcher() - if err != nil { - return err + if _, err := daemon.SdNotify(false, "READY=1"); err != nil { + log.Printf("Error notifying init daemon: %v", err) } - return nil + go checkSqueueForOrphans(dispatcher, sqCheck) + + return dispatcher.Run(context.Background()) +} + +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 checkSqueueForOrphans(dispatcher *dispatch.Dispatcher, sqCheck *SqueueChecker) { + for _, uuid := range sqCheck.All() { + if !containerUuidPattern.MatchString(uuid) { + continue + } + err := dispatcher.TrackContainer(uuid) + if err != nil { + log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err) + } + } } // sbatchCmd -func sbatchFunc(container dispatch.Container) *exec.Cmd { - memPerCPU := math.Ceil((float64(container.RuntimeConstraints["ram"])) / (float64(container.RuntimeConstraints["vcpus"] * 1048576))) - return exec.Command("sbatch", "--share", "--parsable", - fmt.Sprintf("--job-name=%s", container.UUID), - fmt.Sprintf("--mem-per-cpu=%d", int(memPerCPU)), - fmt.Sprintf("--cpus-per-task=%d", int(container.RuntimeConstraints["vcpus"])), - fmt.Sprintf("--priority=%d", container.Priority)) +func sbatchFunc(container arvados.Container) *exec.Cmd { + mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM) / float64(1048576))) + + var disk int64 + for _, m := range container.Mounts { + if m.Kind == "tmp" { + disk += m.Capacity + } + } + disk = int64(math.Ceil(float64(disk) / float64(1048576))) + + var sbatchArgs []string + sbatchArgs = append(sbatchArgs, theConfig.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)) + if len(container.SchedulingParameters.Partitions) > 0 { + sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ","))) + } + + return exec.Command("sbatch", sbatchArgs...) } // scancelCmd -func scancelFunc(container dispatch.Container) *exec.Cmd { +func scancelFunc(container arvados.Container) *exec.Cmd { return exec.Command("scancel", "--name="+container.UUID) } @@ -88,186 +186,122 @@ var sbatchCmd = sbatchFunc var scancelCmd = scancelFunc // Submit job to slurm using sbatch. -func submit(dispatcher *dispatch.Dispatcher, - container dispatch.Container, crunchRunCommand string) (jobid string, submitErr error) { - submitErr = nil - - 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 - } - err := dispatcher.Arv.Update("containers", container.UUID, - arvadosclient.Dict{ - "container": arvadosclient.Dict{"state": "Queued"}}, - nil) - if err != nil { - log.Printf("Error unlocking container %s: %v", container.UUID, err) - } - }() - - // Create the command and attach to stdin/stdout +func submit(dispatcher *dispatch.Dispatcher, container arvados.Container, crunchRunCommand []string) error { 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 - } + // 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))) - stderrReader, stderrErr := cmd.StderrPipe() - if stderrErr != nil { - submitErr = fmt.Errorf("Error creating stderr pipe %v: %q", container.UUID, stderrErr) - return - } + var stdout, stderr bytes.Buffer + cmd.Stdout = &stdout + cmd.Stderr = &stderr // Mutex between squeue sync and running sbatch or scancel. - squeueUpdater.SlurmLock.Lock() - defer squeueUpdater.SlurmLock.Unlock() + sqCheck.L.Lock() + defer sqCheck.L.Unlock() - err := cmd.Start() - if err != nil { - submitErr = fmt.Errorf("Error starting %v: %v", cmd.Args, err) - return - } - - stdoutChan := make(chan []byte) - go func() { - b, _ := ioutil.ReadAll(stdoutReader) - stdoutReader.Close() - stdoutChan <- b - }() - - stderrChan := make(chan []byte) - go func() { - b, _ := ioutil.ReadAll(stderrReader) - stderrReader.Close() - stderrChan <- b - }() - - // Send a tiny script on stdin to execute the crunch-run command - // slurm actually enforces that this must be a #! script - fmt.Fprintf(stdinWriter, "#!/bin/sh\nexec '%s' '%s'\n", crunchRunCommand, container.UUID) - stdinWriter.Close() + log.Printf("exec sbatch %+q", cmd.Args) + err := cmd.Run() - err = cmd.Wait() + switch err.(type) { + case nil: + log.Printf("sbatch succeeded: %q", strings.TrimSpace(stdout.String())) + return nil - stdoutMsg := <-stdoutChan - stderrmsg := <-stderrChan + case *exec.ExitError: + dispatcher.Unlock(container.UUID) + return fmt.Errorf("sbatch %+q failed: %v (stderr: %q)", cmd.Args, err, stderr.Bytes()) - close(stdoutChan) - close(stderrChan) - - if err != nil { - submitErr = fmt.Errorf("Container submission failed %v: %v %v", cmd.Args, err, stderrmsg) - return + default: + dispatcher.Unlock(container.UUID) + return fmt.Errorf("exec failed: %v", err) } +} - // If everything worked out, got the jobid on stdout - jobid = strings.TrimSpace(string(stdoutMsg)) +// 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) { + ctx, cancel := context.WithCancel(context.Background()) + defer cancel() + + if ctr.State == dispatch.Locked && !sqCheck.HasUUID(ctr.UUID) { + log.Printf("Submitting container %s to slurm", ctr.UUID) + if err := submit(disp, ctr, theConfig.CrunchRunCommand); err != nil { + 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 + } + } - 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 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 dispatch.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, *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.UpdateState(container.UUID, dispatch.Queued) - } - 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 dispatch.Container - err := dispatcher.Arv.Get("containers", container.UUID, nil, &con) - if err != nil { - log.Printf("Error getting final container state: %v", err) + // 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) { + } + 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) } - - var st string - switch con.State { - case dispatch.Locked: - st = dispatch.Queued + switch ctr.State { case dispatch.Running: - st = dispatch.Cancelled - 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("Dispatcher says container %s is done: cancel slurm job", ctr.UUID) + 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 in state %v but missing from slurm queue, changing to %v.", - container.UUID, con.State, st) - dispatcher.UpdateState(container.UUID, st) } } } -// 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 dispatch.Container, - status chan dispatch.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() - err := scancelCmd(container).Run() - squeueUpdater.SlurmLock.Unlock() - - if err != nil { - log.Printf("Error stopping container %s with scancel: %v", - container.UUID, err) - 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 scancel(ctr arvados.Container) { + sqCheck.L.Lock() + cmd := scancelCmd(ctr) + msg, err := cmd.CombinedOutput() + sqCheck.L.Unlock() + + if err != nil { + log.Printf("%q %q: %s %q", cmd.Path, cmd.Args, err, msg) + time.Sleep(time.Second) + } else if 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 } - monitorDone = true + return err }