Arvados-DCO-1.1-Signed-off-by: Radhika Chippada <radhika@curoverse.com>
[arvados.git] / services / crunch-dispatch-slurm / crunch-dispatch-slurm.go
index 617b076da281a982b81f24ae9b7b7fe4d3897aee..30cbb79dc186de45366d648e2a26e42266ee7de3 100644 (file)
@@ -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.
@@ -11,6 +15,7 @@ import (
        "math"
        "os"
        "os/exec"
+       "regexp"
        "strings"
        "time"
 
@@ -122,19 +127,48 @@ func doMain() error {
                log.Printf("Error notifying init daemon: %v", err)
        }
 
+       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))
+       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, ",")))
        }
@@ -196,13 +230,21 @@ 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