X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/83c8f1685d812c31d8bd568f3c2ac1edcd120aed..ca06cfbda0e84d469f7810a280cfa4dfa8997260:/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..e743fa5854 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.
@@ -19,6 +23,7 @@ import (
 	"git.curoverse.com/arvados.git/sdk/go/arvadosclient"
 	"git.curoverse.com/arvados.git/sdk/go/config"
 	"git.curoverse.com/arvados.git/sdk/go/dispatch"
+	arvadosVersion "git.curoverse.com/arvados.git/sdk/go/version"
 	"github.com/coreos/go-systemd/daemon"
 )
 
@@ -65,10 +70,21 @@ func doMain() error {
 		"dump-config",
 		false,
 		"write current configuration to stdout and exit")
-
+	getVersion := flags.Bool(
+		"version",
+		false,
+		"Print version information and exit.")
 	// Parse args; omit the first arg which is the command name
 	flags.Parse(os.Args[1:])
 
+	// Print version information if requested
+	if *getVersion {
+		fmt.Printf("Version: %s\n", arvadosVersion.GetVersion())
+		os.Exit(0)
+	}
+
+	log.Printf("crunch-dispatch-slurm %q started", arvadosVersion.GetVersion())
+
 	err := readConfig(&theConfig, *configPath)
 	if err != nil {
 		return err
@@ -123,44 +139,48 @@ 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)
+		}
+	}
 }
 
 // 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))
 	if len(container.SchedulingParameters.Partitions) > 0 {
 		sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ",")))
 	}
@@ -222,13 +242,21 @@ 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