X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/e39e6831a61af3b889a78b2380bf978c48a20fb9..18de85806e3717421accb89b093fc2bd56822100:/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 e053c21d7b..aaea51cbf7 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -1,186 +1,328 @@ package main +// Dispatcher service for Crunch that submits containers to the slurm queue. + import ( "flag" "fmt" + "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" "io" + "io/ioutil" "log" + "math" "os" "os/exec" - "os/signal" - "sync" - "syscall" + "strings" "time" ) +// 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 +} + 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 + squeueUpdater Squeue ) +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") + configPath := flags.String( + "config", + defaultConfigPath, + "`path` to JSON or YAML configuration file") // 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"} + } - // 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.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_INSECURE", "") + if theConfig.Client.Insecure { + os.Setenv("ARVADOS_API_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) + arv, err := arvadosclient.MakeArvadosClient() + if err != nil { + log.Printf("Error making Arvados client: %v", err) + return err + } + arv.Retries = 25 - // Wait for all running crunch jobs to complete / terminate - waitGroup.Wait() + squeueUpdater.StartMonitor(time.Duration(theConfig.PollPeriod)) + defer squeueUpdater.Done() - return nil -} + dispatcher := dispatch.Dispatcher{ + Arv: arv, + RunContainer: run, + PollInterval: time.Duration(theConfig.PollPeriod), + DoneProcessing: make(chan struct{})} -// Poll for queued containers using pollInterval. -// Invoke dispatchLocal 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 string) { - ticker := time.NewTicker(time.Duration(pollInterval) * time.Second) - - for { - select { - case <-ticker.C: - dispatchSlurm(priorityPollInterval, crunchRunCommand) - case <-doneProcessing: - ticker.Stop() - return - } + if _, err := daemon.SdNotify("READY=1"); err != nil { + log.Printf("Error notifying init daemon: %v", err) + } + + err = dispatcher.RunDispatcher() + if err != nil { + return err } + + return nil } -// Container data -type Container struct { - UUID string `json:"uuid"` - State string `json:"state"` - Priority int `json:"priority"` +// sbatchCmd +func sbatchFunc(container arvados.Container) *exec.Cmd { + memPerCPU := math.Ceil(float64(container.RuntimeConstraints.RAM) / (float64(container.RuntimeConstraints.VCPUs) * 1048576)) + + 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)) + + return exec.Command("sbatch", sbatchArgs...) } -// ContainerList is a list of the containers from api -type ContainerList struct { - Items []Container `json:"items"` +// scancelCmd +func scancelFunc(container arvados.Container) *exec.Cmd { + return exec.Command("scancel", "--name="+container.UUID) } -// Get the list of queued containers from API server and invoke run for each container. -func dispatchSlurm(priorityPollInterval int, crunchRunCommand string) { - params := arvadosclient.Dict{ - "filters": [][]string{[]string{"state", "=", "Queued"}}, +// 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 + } + err := dispatcher.Unlock(container.UUID) + if err != nil { + log.Printf("Error unlocking container %s: %v", container.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 } - var containers ContainerList - err := arv.List("containers", params, &containers) - if err != nil { - log.Printf("Error getting list of queued containers: %q", err) + stdoutReader, stdoutErr := cmd.StdoutPipe() + if stdoutErr != nil { + submitErr = fmt.Errorf("Error creating stdout pipe %v: %q", container.UUID, stdoutErr) return } - 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].UUID, crunchRunCommand, priorityPollInterval) + stderrReader, stderrErr := cmd.StderrPipe() + if stderrErr != nil { + submitErr = fmt.Errorf("Error creating stderr pipe %v: %q", container.UUID, stderrErr) + return } -} -// Run queued container: -// Set container state to locked (TBD) -// Run container using the given crunch-run command -// Set the container state to Running -// If the container priority becomes zero while crunch job is still running, terminate it. -func run(uuid string, crunchRunCommand string, priorityPollInterval int) { - stdinReader, stdinWriter := io.Pipe() - - cmd := exec.Command("sbatch", "--job-name="+uuid) - cmd.Stdin = stdinReader - cmd.Stderr = os.Stderr - cmd.Stdout = os.Stderr - if err := cmd.Start(); err != nil { - log.Printf("Error running container for %v: %q", uuid, err) + // Mutex between squeue sync and running sbatch or scancel. + squeueUpdater.SlurmLock.Lock() + defer squeueUpdater.SlurmLock.Unlock() + + err := cmd.Start() + if err != nil { + submitErr = fmt.Errorf("Error starting %v: %v", cmd.Args, err) return } - fmt.Fprintf(stdinWriter, "#!/bin/sh\nexec %s %s\n", crunchRunCommand, uuid) + 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 + io.WriteString(stdinWriter, execScript(append(crunchRunCommand, container.UUID))) stdinWriter.Close() - cmd.Wait() - // Update container status to Running - err := arv.Update("containers", uuid, - arvadosclient.Dict{ - "container": arvadosclient.Dict{"state": "Running"}}, - nil) + err = cmd.Wait() + + stdoutMsg := <-stdoutChan + stderrmsg := <-stderrChan + + close(stdoutChan) + close(stderrChan) + if err != nil { - log.Printf("Error updating container state to 'Running' for %v: %q", uuid, err) + submitErr = fmt.Errorf("Container submission failed: %v: %v (stderr: %q)", cmd.Args, err, stderrmsg) + return } - log.Printf("Submitted container run for %v", uuid) + log.Printf("sbatch succeeded: %s", strings.TrimSpace(string(stdoutMsg))) + return +} - // 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", uuid, nil, &container) +// 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 container info for %v: %q", uuid, err) - } else { - if container.Priority == 0 { - log.Printf("Canceling container %v", uuid) - priorityTicker.Stop() - cancelcmd := exec.Command("scancel", "--name="+uuid) - cancelcmd.Run() + log.Printf("Error getting final container state: %v", err) + } + + 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) + 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 + } + } + } +} + +// 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() + 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) } } - }() + } + monitorDone = true +} +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 + } + return err }