+// 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"
"log"
"math"
"os"
"os/exec"
+ "regexp"
"strings"
"time"
var (
theConfig Config
- sqCheck SqueueChecker
+ sqCheck = &SqueueChecker{}
)
const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
}
arv.Retries = 25
- sqCheck = SqueueChecker{Period: time.Duration(theConfig.PollPeriod)}
+ sqCheck = &SqueueChecker{Period: time.Duration(theConfig.PollPeriod)}
defer sqCheck.Stop()
- dispatcher := dispatch.Dispatcher{
+ dispatcher := &dispatch.Dispatcher{
Arv: arv,
RunContainer: run,
PollPeriod: time.Duration(theConfig.PollPeriod),
log.Printf("Error notifying init daemon: %v", err)
}
- return dispatcher.Run()
+ 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 arvados.Container) *exec.Cmd {
- memPerCPU := math.Ceil(float64(container.RuntimeConstraints.RAM) / (float64(container.RuntimeConstraints.VCPUs) * 1048576))
+ 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, "--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("--mem=%d", mem))
sbatchArgs = append(sbatchArgs, fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs))
- if container.SchedulingParameters.Partitions != nil {
+ 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, ",")))
}
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
- }
- dispatcher.Unlock(container.UUID)
- }()
-
+func submit(dispatcher *dispatch.Dispatcher, container arvados.Container, crunchRunCommand []string) error {
cmd := sbatchCmd(container)
// Send a tiny script on stdin to execute the crunch-run
log.Printf("exec sbatch %+q", cmd.Args)
err := cmd.Run()
+
switch err.(type) {
case nil:
log.Printf("sbatch succeeded: %q", strings.TrimSpace(stdout.String()))
return nil
+
case *exec.ExitError:
+ dispatcher.Unlock(container.UUID)
return fmt.Errorf("sbatch %+q failed: %v (stderr: %q)", cmd.Args, err, stderr.Bytes())
+
default:
+ dispatcher.Unlock(container.UUID)
return fmt.Errorf("exec failed: %v", err)
}
}
-// 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 sqCheck.HasUUID(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)
- }
+// 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
+ }
+ }
- 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)
+ 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 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)
+ }
+ switch ctr.State {
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
+ 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 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.Priority == 0 && (container.State == dispatch.Locked || container.State == dispatch.Running) {
- log.Printf("Canceling container %s", container.UUID)
- // Mutex between squeue sync and running sbatch or scancel.
- sqCheck.L.Lock()
- cmd := scancelCmd(container)
- msg, err := cmd.CombinedOutput()
- sqCheck.L.Unlock()
-
- if err != nil {
- log.Printf("Error stopping container %s with %v %v: %v %v", container.UUID, cmd.Path, cmd.Args, err, string(msg))
- if sqCheck.HasUUID(container.UUID) {
- log.Printf("Container %s is still in squeue after scancel.", container.UUID)
- continue
- }
- }
+func scancel(ctr arvados.Container) {
+ sqCheck.L.Lock()
+ cmd := scancelCmd(ctr)
+ msg, err := cmd.CombinedOutput()
+ sqCheck.L.Unlock()
- // Ignore errors; if necessary, we'll try again next time
- dispatcher.UpdateState(container.UUID, dispatch.Cancelled)
- }
+ 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)
}
- monitorDone = true
}
func readConfig(dst interface{}, path string) error {