X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/1695ccb6279de083eeed80d36d38f607a5c7098d..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 07653b1de1..30cbb79dc1 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -1,305 +1,307 @@ +// 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" - "io/ioutil" "log" + "math" "os" "os/exec" - "os/signal" - "strconv" - "sync" - "syscall" + "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 ( - arv arvadosclient.ArvadosClient - runningCmds map[string]*exec.Cmd - runningCmdsMutex sync.Mutex - waitGroup sync.WaitGroup - doneProcessing chan bool - sigChan chan os.Signal + 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") - - priorityPollInterval := flags.Int( - "container-priority-poll-interval", - 60, - "Interval in seconds to check priority of a dispatched container") - - crunchRunCommand := flags.String( - "crunch-run-command", - "/usr/bin/crunch-run", - "Crunch command to run container") - - finishCommand := flags.String( - "finish-command", - "/usr/bin/crunch-finish-slurm.sh", - "Command to run from strigger when job is finished") + 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:]) - var err error - arv, err = arvadosclient.MakeArvadosClient() + err := readConfig(&theConfig, *configPath) if err != nil { return err } - // Channel to terminate - doneProcessing = make(chan bool) + if theConfig.CrunchRunCommand == nil { + theConfig.CrunchRunCommand = []string{"crunch-run"} + } + + if theConfig.PollPeriod == 0 { + theConfig.PollPeriod = arvados.Duration(10 * time.Second) + } - // Graceful shutdown - sigChan = make(chan os.Signal, 1) - signal.Notify(sigChan, syscall.SIGINT, syscall.SIGTERM, syscall.SIGQUIT) - go func(sig <-chan os.Signal) { - for sig := range sig { - log.Printf("Caught signal: %v", sig) - doneProcessing <- true + 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") } - }(sigChan) + 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).") + } - // Run all queued containers - runQueuedContainers(*pollInterval, *priorityPollInterval, *crunchRunCommand, *finishCommand) + if *dumpConfig { + log.Fatal(config.DumpAndExit(theConfig)) + } - // Wait for all running crunch jobs to complete / terminate - waitGroup.Wait() + arv, err := arvadosclient.MakeArvadosClient() + if err != nil { + log.Printf("Error making Arvados client: %v", err) + return err + } + arv.Retries = 25 - return nil -} + sqCheck = &SqueueChecker{Period: time.Duration(theConfig.PollPeriod)} + defer sqCheck.Stop() -// Poll for queued containers using pollInterval. -// Invoke dispatchSlurm for each ticker cycle, which will run all the queued containers. -// -// Any errors encountered are logged but the program would continue to run (not exit). -// This is because, once one or more crunch jobs are running, -// we would need to wait for them complete. -func runQueuedContainers(pollInterval, priorityPollInterval int, crunchRunCommand, finishCommand string) { - ticker := time.NewTicker(time.Duration(pollInterval) * time.Second) + dispatcher := &dispatch.Dispatcher{ + Arv: arv, + RunContainer: run, + PollPeriod: time.Duration(theConfig.PollPeriod), + MinRetryPeriod: time.Duration(theConfig.MinRetryPeriod), + } - for { - select { - case <-ticker.C: - dispatchSlurm(priorityPollInterval, crunchRunCommand, finishCommand) - case <-doneProcessing: - ticker.Stop() - return - } + if _, err := daemon.SdNotify(false, "READY=1"); err != nil { + log.Printf("Error notifying init daemon: %v", err) } -} -// Container data -type Container struct { - UUID string `json:"uuid"` - State string `json:"state"` - Priority int `json:"priority"` - RuntimeConstraints map[string]int `json:"runtime_constraints"` -} + go checkSqueueForOrphans(dispatcher, sqCheck) -// ContainerList is a list of the containers from api -type ContainerList struct { - Items []Container `json:"items"` + return dispatcher.Run(context.Background()) } -// Get the list of queued containers from API server and invoke run for each container. -func dispatchSlurm(priorityPollInterval int, crunchRunCommand, finishCommand string) { - params := arvadosclient.Dict{ - "filters": [][]string{[]string{"state", "=", "Queued"}}, +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) + } } +} - var containers ContainerList - err := arv.List("containers", params, &containers) - if err != nil { - log.Printf("Error getting list of queued containers: %q", err) - return - } +// sbatchCmd +func sbatchFunc(container arvados.Container) *exec.Cmd { + mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM) / float64(1048576))) - for i := 0; i < len(containers.Items); i++ { - log.Printf("About to submit queued container %v", containers.Items[i].UUID) - // Run the container - go run(containers.Items[i], crunchRunCommand, finishCommand, priorityPollInterval) + 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, ","))) } -} -// sbatchCmd -func sbatchFunc(container Container) *exec.Cmd { - return exec.Command("sbatch", "--share", "--parsable", - "--job-name="+container.UUID, - "--mem="+strconv.Itoa(container.RuntimeConstraints["ram"]), - "--cpus-per-task="+strconv.Itoa(container.RuntimeConstraints["vcpus"])) + return exec.Command("sbatch", sbatchArgs...) } -var sbatchCmd = sbatchFunc - -// striggerCmd -func striggerFunc(jobid, containerUUID, finishCommand, apiHost, apiToken, apiInsecure string) *exec.Cmd { - return exec.Command("strigger", "--set", "--jobid="+jobid, "--fini", - fmt.Sprintf("--program=%s %s %s %s %s", finishCommand, apiHost, apiToken, apiInsecure, containerUUID)) +// scancelCmd +func scancelFunc(container arvados.Container) *exec.Cmd { + return exec.Command("scancel", "--name="+container.UUID) } -var striggerCmd = striggerFunc +// Wrap these so that they can be overridden by tests +var sbatchCmd = sbatchFunc +var scancelCmd = scancelFunc // Submit job to slurm using sbatch. -func submit(container Container, crunchRunCommand string) (jobid string, submitErr error) { - submitErr = nil - - // Mark record as complete if anything errors out. - defer func() { - if submitErr != nil { - // This really should be an "Error" state, see #8018 - updateErr := arv.Update("containers", container.UUID, - arvadosclient.Dict{ - "container": arvadosclient.Dict{"state": "Complete"}}, - nil) - if updateErr != nil { - log.Printf("Error updating container state to 'Complete' for %v: %q", container.UUID, updateErr) - } - } - }() - - // 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 - } - - stderrReader, stderrErr := cmd.StderrPipe() - if stderrErr != nil { - submitErr = fmt.Errorf("Error creating stderr pipe %v: %q", container.UUID, stderrErr) - 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))) - err := cmd.Start() - if err != nil { - submitErr = fmt.Errorf("Error starting %v: %v", cmd.Args, err) - return - } + var stdout, stderr bytes.Buffer + cmd.Stdout = &stdout + cmd.Stderr = &stderr - stdoutChan := make(chan []byte) - go func() { - b, _ := ioutil.ReadAll(stdoutReader) - stdoutChan <- b - close(stdoutChan) - }() + // Mutex between squeue sync and running sbatch or scancel. + sqCheck.L.Lock() + defer sqCheck.L.Unlock() - stderrChan := make(chan []byte) - go func() { - b, _ := ioutil.ReadAll(stderrReader) - stderrChan <- b - close(stderrChan) - }() + log.Printf("exec sbatch %+q", cmd.Args) + err := cmd.Run() - // 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() + switch err.(type) { + case nil: + log.Printf("sbatch succeeded: %q", strings.TrimSpace(stdout.String())) + return nil - err = cmd.Wait() + case *exec.ExitError: + dispatcher.Unlock(container.UUID) + return fmt.Errorf("sbatch %+q failed: %v (stderr: %q)", cmd.Args, err, stderr.Bytes()) - stdoutMsg := <-stdoutChan - stderrmsg := <-stderrChan + default: + dispatcher.Unlock(container.UUID) + return fmt.Errorf("exec failed: %v", err) + } +} - if err != nil { - submitErr = fmt.Errorf("Container submission failed %v: %v %v", cmd.Args, err, stderrmsg) - return +// 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 + } } - // If everything worked out, got the jobid on stdout - jobid = string(stdoutMsg) + log.Printf("Start monitoring container %v in state %q", ctr.UUID, ctr.State) + defer log.Printf("Done monitoring container %s", ctr.UUID) - return -} + // 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) -// finalizeRecordOnFinish uses 'strigger' command to register a script that will run on -// the slurm controller when the job finishes. -func finalizeRecordOnFinish(jobid, containerUUID, finishCommand, apiHost, apiToken, apiInsecure string) { - cmd := striggerCmd(jobid, containerUUID, finishCommand, apiHost, apiToken, apiInsecure) - cmd.Stdout = os.Stdout - cmd.Stderr = os.Stderr - err := cmd.Run() - if err != nil { - log.Printf("While setting up strigger: %v", err) + 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: + 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) + } + } } } -// Run a queued container. -// Set container state to locked (TBD) -// Submit job to slurm to execute crunch-run command for the container -// If the container priority becomes zero while crunch job is still running, cancel the job. -func run(container Container, crunchRunCommand, finishCommand string, priorityPollInterval int) { +func scancel(ctr arvados.Container) { + sqCheck.L.Lock() + cmd := scancelCmd(ctr) + msg, err := cmd.CombinedOutput() + sqCheck.L.Unlock() - jobid, err := submit(container, crunchRunCommand) if err != nil { - log.Printf("Error queuing container run: %v", err) - return + 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) } +} - insecure := "0" - if arv.ApiInsecure { - insecure = "1" - } - finalizeRecordOnFinish(jobid, container.UUID, finishCommand, arv.ApiServer, arv.ApiToken, insecure) - - // Update container status to Running, this is a temporary workaround - // to avoid resubmitting queued containers because record locking isn't - // implemented yet. - err = arv.Update("containers", container.UUID, - arvadosclient.Dict{ - "container": arvadosclient.Dict{"state": "Running"}}, - nil) - if err != nil { - log.Printf("Error updating container state to 'Running' for %v: %q", container.UUID, err) +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 } - - log.Printf("Submitted container run for %v", container.UUID) - - containerUUID := container.UUID - - // A goroutine to terminate the runner if container priority becomes zero - priorityTicker := time.NewTicker(time.Duration(priorityPollInterval) * time.Second) - go func() { - for _ = range priorityTicker.C { - var container Container - err := arv.Get("containers", containerUUID, nil, &container) - if err != nil { - log.Printf("Error getting container info for %v: %q", container.UUID, err) - } else { - if container.Priority == 0 { - log.Printf("Canceling container %v", container.UUID) - priorityTicker.Stop() - cancelcmd := exec.Command("scancel", "--name="+container.UUID) - cancelcmd.Run() - } - if container.State == "Complete" { - priorityTicker.Stop() - } - } - } - }() - + return err }