package main
+// Dispatcher service for Crunch that submits containers to the slurm queue.
+
import (
"flag"
"fmt"
"git.curoverse.com/arvados.git/sdk/go/arvadosclient"
- "io"
+ "git.curoverse.com/arvados.git/sdk/go/dispatch"
+ "io/ioutil"
"log"
+ "math"
"os"
"os/exec"
- "os/signal"
- "sync"
- "syscall"
+ "strings"
"time"
)
}
var (
- arv arvadosclient.ArvadosClient
- runningCmds map[string]*exec.Cmd
- runningCmdsMutex sync.Mutex
- waitGroup sync.WaitGroup
- doneProcessing chan bool
- sigChan chan os.Signal
+ crunchRunCommand *string
+ squeueUpdater Squeue
)
func doMain() error {
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(
+ crunchRunCommand = flags.String(
"crunch-run-command",
"/usr/bin/crunch-run",
"Crunch command to run container")
// Parse args; omit the first arg which is the command name
flags.Parse(os.Args[1:])
- var err error
- arv, err = arvadosclient.MakeArvadosClient()
+ arv, err := arvadosclient.MakeArvadosClient()
if err != nil {
+ log.Printf("Error making Arvados client: %v", err)
return err
}
+ arv.Retries = 25
- // Channel to terminate
- doneProcessing = make(chan bool)
-
- // 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
- }
- }(sigChan)
+ squeueUpdater.StartMonitor(time.Duration(*pollInterval) * time.Second)
+ defer squeueUpdater.Done()
- // Run all queued containers
- runQueuedContainers(*pollInterval, *priorityPollInterval, *crunchRunCommand)
+ dispatcher := dispatch.Dispatcher{
+ Arv: arv,
+ RunContainer: run,
+ PollInterval: time.Duration(*pollInterval) * time.Second,
+ DoneProcessing: make(chan struct{})}
- // Wait for all running crunch jobs to complete / terminate
- waitGroup.Wait()
+ err = dispatcher.RunDispatcher()
+ if err != nil {
+ return err
+ }
return nil
}
-// 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
- }
- }
+// 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))
}
-// Container data
-type Container struct {
- UUID string `json:"uuid"`
- State string `json:"state"`
- Priority int `json:"priority"`
+// scancelCmd
+func scancelFunc(container dispatch.Container) *exec.Cmd {
+ return exec.Command("scancel", "--name="+container.UUID)
}
-// ContainerList is a list of the containers from api
-type ContainerList struct {
- Items []Container `json:"items"`
-}
+// 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 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)
+ }
+ }()
-// 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"}},
+ // 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
+ fmt.Fprintf(stdinWriter, "#!/bin/sh\nexec '%s' '%s'\n", 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 %v", cmd.Args, err, stderrmsg)
+ return
}
- log.Printf("Submitted container run for %v", uuid)
+ // If everything worked out, got the jobid on stdout
+ jobid = strings.TrimSpace(string(stdoutMsg))
- // 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)
+ return
+}
+
+// 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 container info for %v: %q", uuid, err)
- } else {
- if container.Priority == 0 {
- priorityTicker.Stop()
- cancelcmd := exec.Command("scancel", "--name="+uuid)
- cancelcmd.Run()
- }
+ log.Printf("Error getting final container state: %v", err)
}
+
+ var st string
+ switch con.State {
+ case dispatch.Locked:
+ st = dispatch.Queued
+ case dispatch.Running:
+ st = dispatch.Cancelled
+ default:
+ // Container state is Queued, Complete or Cancelled so stop monitoring it.
+ return
+ }
+
+ 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)
+ }
+ }
+ }
+ monitorDone = true
}
projects / arvados.git / blobdiff