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")
-
- 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")
// 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_INSECURE", "")
+ if theConfig.Client.Insecure {
+ os.Setenv("ARVADOS_API_INSECURE", "1")
}
- }(sigChan)
+ os.Setenv("ARVADOS_KEEP_SERVICES", "")
+ 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)
+ 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, finishCommand string) {
- ticker := time.NewTicker(time.Duration(pollInterval) * time.Second)
-
- for {
- select {
- case <-ticker.C:
- dispatchSlurm(priorityPollInterval, crunchRunCommand, finishCommand)
- case <-doneProcessing:
- ticker.Stop()
- return
- }
+ if _, err := daemon.SdNotify("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"`
-}
+ err = dispatcher.RunDispatcher()
+ if err != nil {
+ return err
+ }
-// ContainerList is a list of the containers from api
-type ContainerList struct {
- Items []Container `json:"items"`
+ return nil
}
-// 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"}},
- }
+// sbatchCmd
+func sbatchFunc(container arvados.Container) *exec.Cmd {
+ memPerCPU := math.Ceil(float64(container.RuntimeConstraints.RAM) / (float64(container.RuntimeConstraints.VCPUs) * 1048576))
- var containers ContainerList
- err := arv.List("containers", params, &containers)
- if err != nil {
- log.Printf("Error getting list of queued containers: %q", err)
- return
- }
+ 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))
- 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)
- }
+ return exec.Command("sbatch", sbatchArgs...)
+}
+
+// scancelCmd
+func scancelFunc(container arvados.Container) *exec.Cmd {
+ return exec.Command("scancel", "--name="+container.UUID)
}
-func submit(container Container, crunchRunCommand string) (jobid string, submiterr error) {
- submiterr = nil
+// 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 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)
- }
+ // 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)
}
}()
- cmd := exec.Command("sbatch", "--job-name="+container.UUID, "--share", "--parsable")
+ // 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)
+ 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)
+ 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)
+ stderrReader, stderrErr := cmd.StderrPipe()
+ if stderrErr != nil {
+ submitErr = fmt.Errorf("Error creating stderr pipe %v: %q", container.UUID, stderrErr)
return
}
+ // 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)
+ submitErr = fmt.Errorf("Error starting %v: %v", cmd.Args, err)
return
}
- stdoutchan := make(chan []byte)
+ stdoutChan := make(chan []byte)
go func() {
b, _ := ioutil.ReadAll(stdoutReader)
- stdoutchan <- b
- close(stdoutchan)
+ stdoutReader.Close()
+ stdoutChan <- b
}()
- stderrchan := make(chan []byte)
+ stderrChan := make(chan []byte)
go func() {
b, _ := ioutil.ReadAll(stderrReader)
- stderrchan <- b
- close(stderrchan)
+ stderrReader.Close()
+ stderrChan <- b
}()
- fmt.Fprintf(stdinWriter, "#!/bin/sh\nexec '%s' '%s'\n", crunchRunCommand, container.UUID)
+ // 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()
err = cmd.Wait()
- stdoutmsg := <-stdoutchan
- stderrmsg := <-stderrchan
+ stdoutMsg := <-stdoutChan
+ stderrmsg := <-stderrChan
+
+ close(stdoutChan)
+ close(stderrChan)
if err != nil {
- submiterr = fmt.Errorf("Container submission failed %v: %v %v", cmd.Args, err, stderrmsg)
+ submitErr = fmt.Errorf("Container submission failed: %v: %v (stderr: %q)", cmd.Args, err, stderrmsg)
return
}
- jobid = string(stdoutmsg)
-
+ log.Printf("sbatch succeeded: %s", strings.TrimSpace(string(stdoutMsg)))
return
}
-func strigger(jobid, containerUUID, finishCommand string) {
- cmd := exec.Command("strigger", "--set", "--jobid="+jobid, "--fini", fmt.Sprintf("--program=%s", finishCommand))
- cmd.Stdout = os.Stdout
- cmd.Stderr = os.Stderr
- err := cmd.Run()
- if err != nil {
- log.Printf("While setting up strigger: %v", err)
- }
-}
-
-// 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(container Container, crunchRunCommand, finishCommand string, priorityPollInterval int) {
-
- jobid, err := submit(container, crunchRunCommand)
- if err != nil {
- log.Printf("Error queuing container run: %v", err)
- return
- }
-
- strigger(jobid, container.UUID, finishCommand)
+// 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)
+ }
- // Update container status to Running
- 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)
+ 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
+ }
+ }
}
+}
- 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()
+// 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
}