X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/353eb4f4c4fb52e7f2a1c9aaad93e9d6bf0088f4..0bc641e077a050b3f86957514fae0e4d8ba1770b:/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 7cb14fab15..2f30ebfda2 100644
--- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go
+++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go
@@ -1,3 +1,7 @@
+// 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.
@@ -123,44 +127,59 @@ func doMain() error {
 		log.Printf("Error notifying init daemon: %v", err)
 	}
 
-	containerTrackerTicker := trackContainers(dispatcher)
-	defer containerTrackerTicker.Stop()
+	go checkSqueueForOrphans(dispatcher, sqCheck)
 
 	return dispatcher.Run(context.Background())
 }
 
-var containerUuidPattern = regexp.MustCompile(`[a-z0-9]{5}-dz642-[a-z0-9]{15}$`)
-
-// Start a goroutine to check squeue report periodically, and
-// invoke TrackContainer for all the containers in the report.
-func trackContainers(dispatcher *dispatch.Dispatcher) *time.Ticker {
-	ticker := time.NewTicker(sqCheck.Period)
-	go func() {
-		for {
-			select {
-			case <-ticker.C:
-				for uuid := range sqCheck.AllUuids() {
-					match := containerUuidPattern.MatchString(uuid)
-					if match {
-						dispatcher.TrackContainer(uuid)
-					}
-				}
-			}
+var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$`)
+
+// 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
 		}
-	}()
-	return ticker
+		err := dispatcher.TrackContainer(uuid)
+		if err != nil {
+			log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err)
+		}
+	}
+}
+
+func niceness(priority int) int {
+	if priority > 1000 {
+		priority = 1000
+	}
+	if priority < 0 {
+		priority = 0
+	}
+	return (1000 - priority) * 1000
 }
 
 // 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))
+	sbatchArgs = append(sbatchArgs, fmt.Sprintf("--tmp=%d", disk))
+	sbatchArgs = append(sbatchArgs, fmt.Sprintf("--nice=%d", niceness(container.Priority)))
 	if len(container.SchedulingParameters.Partitions) > 0 {
 		sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ",")))
 	}
@@ -173,9 +192,15 @@ func scancelFunc(container arvados.Container) *exec.Cmd {
 	return exec.Command("scancel", "--name="+container.UUID)
 }
 
+// scontrolCmd
+func scontrolFunc(container arvados.Container) *exec.Cmd {
+	return exec.Command("scontrol", "update", "JobName="+container.UUID, fmt.Sprintf("--nice=%d", niceness(container.Priority)))
+}
+
 // Wrap these so that they can be overridden by tests
 var sbatchCmd = sbatchFunc
 var scancelCmd = scancelFunc
+var scontrolCmd = scontrolFunc
 
 // Submit job to slurm using sbatch.
 func submit(dispatcher *dispatch.Dispatcher, container arvados.Container, crunchRunCommand []string) error {
@@ -222,13 +247,21 @@ func run(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados
 	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 {
-			log.Printf("Error submitting container %s to slurm: %s", ctr.UUID, err)
+			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
 		}
 	}
 
-	log.Printf("Start monitoring container %s", ctr.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
@@ -261,6 +294,9 @@ func run(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados
 			} 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)
+			} else if niceness(updated.Priority) != sqCheck.GetNiceness(ctr.UUID) && sqCheck.GetNiceness(ctr.UUID) != -1 {
+				// dynamically adjust priority
+				scontrolUpdate(updated)
 			}
 		}
 	}
@@ -281,6 +317,21 @@ func scancel(ctr arvados.Container) {
 	}
 }
 
+func scontrolUpdate(ctr arvados.Container) {
+	sqCheck.L.Lock()
+	cmd := scontrolCmd(ctr)
+	msg, err := cmd.CombinedOutput()
+	sqCheck.L.Unlock()
+
+	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 priority is now %v, niceness is now %v",
+			ctr.UUID, ctr.Priority, sqCheck.GetNiceness(ctr.UUID))
+	}
+}
+
 func readConfig(dst interface{}, path string) error {
 	err := config.LoadFile(dst, path)
 	if err != nil && os.IsNotExist(err) && path == defaultConfigPath {