lib/dispatchcloud/scheduler/run_queue.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package scheduler
   6
   7 import (
   8         "sort"
   9         "time"
  10
  11         "git.arvados.org/arvados.git/lib/dispatchcloud/container"
  12         "git.arvados.org/arvados.git/sdk/go/arvados"
  13         "github.com/sirupsen/logrus"
  14 )
  15
  16 func (sch *Scheduler) runQueue() {
  17         unsorted, _ := sch.queue.Entries()
  18         sorted := make([]container.QueueEnt, 0, len(unsorted))
  19         for _, ent := range unsorted {
  20                 sorted = append(sorted, ent)
  21         }
  22         sort.Slice(sorted, func(i, j int) bool {
  23                 return sorted[i].Container.Priority > sorted[j].Container.Priority
  24         })
  25
  26         running := sch.pool.Running()
  27         unalloc := sch.pool.Unallocated()
  28
  29         sch.logger.WithFields(logrus.Fields{
  30                 "Containers": len(sorted),
  31                 "Processes":  len(running),
  32         }).Debug("runQueue")
  33
  34         dontstart := map[arvados.InstanceType]bool{}
  35         var overquota []container.QueueEnt // entries that are unmappable because of worker pool quota
  36         var containerAllocatedWorkerBootingCount int
  37
  38 tryrun:
  39         for i, ctr := range sorted {
  40                 ctr, it := ctr.Container, ctr.InstanceType
  41                 logger := sch.logger.WithFields(logrus.Fields{
  42                         "ContainerUUID": ctr.UUID,
  43                         "InstanceType":  it.Name,
  44                 })
  45                 if _, running := running[ctr.UUID]; running || ctr.Priority < 1 {
  46                         continue
  47                 }
  48                 switch ctr.State {
  49                 case arvados.ContainerStateQueued:
  50                         if unalloc[it] < 1 && sch.pool.AtQuota() {
  51                                 logger.Debug("not locking: AtQuota and no unalloc workers")
  52                                 overquota = sorted[i:]
  53                                 break tryrun
  54                         }
  55                         if sch.pool.KillContainer(ctr.UUID, "about to lock") {
  56                                 logger.Info("not locking: crunch-run process from previous attempt has not exited")
  57                                 continue
  58                         }
  59                         go sch.lockContainer(logger, ctr.UUID)
  60                         unalloc[it]--
  61                 case arvados.ContainerStateLocked:
  62                         if unalloc[it] > 0 {
  63                                 unalloc[it]--
  64                         } else if sch.pool.AtQuota() {
  65                                 // Don't let lower-priority containers
  66                                 // starve this one by using keeping
  67                                 // idle workers alive on different
  68                                 // instance types.
  69                                 logger.Debug("unlocking: AtQuota and no unalloc workers")
  70                                 sch.queue.Unlock(ctr.UUID)
  71                                 overquota = sorted[i:]
  72                                 break tryrun
  73                         } else if logger.Info("creating new instance"); sch.pool.Create(it) {
  74                                 // Success. (Note pool.Create works
  75                                 // asynchronously and does its own
  76                                 // logging, so we don't need to.)
  77                         } else {
  78                                 // Failed despite not being at quota,
  79                                 // e.g., cloud ops throttled.  TODO:
  80                                 // avoid getting starved here if
  81                                 // instances of a specific type always
  82                                 // fail.
  83                                 continue
  84                         }
  85
  86                         if dontstart[it] {
  87                                 // We already tried & failed to start
  88                                 // a higher-priority container on the
  89                                 // same instance type. Don't let this
  90                                 // one sneak in ahead of it.
  91                         } else if sch.pool.KillContainer(ctr.UUID, "about to start") {
  92                                 logger.Info("not restarting yet: crunch-run process from previous attempt has not exited")
  93                         } else if sch.pool.StartContainer(it, ctr) {
  94                                 // Success.
  95                         } else {
  96                                 containerAllocatedWorkerBootingCount += 1
  97                                 dontstart[it] = true
  98                         }
  99                 }
 100         }
 101
 102         sch.mContainersAllocatedNotStarted.Set(float64(containerAllocatedWorkerBootingCount))
 103         sch.mContainersNotAllocatedOverQuota.Set(float64(len(overquota)))
 104
 105         if len(overquota) > 0 {
 106                 // Unlock any containers that are unmappable while
 107                 // we're at quota.
 108                 for _, ctr := range overquota {
 109                         ctr := ctr.Container
 110                         if ctr.State == arvados.ContainerStateLocked {
 111                                 logger := sch.logger.WithField("ContainerUUID", ctr.UUID)
 112                                 logger.Debug("unlock because pool capacity is used by higher priority containers")
 113                                 err := sch.queue.Unlock(ctr.UUID)
 114                                 if err != nil {
 115                                         logger.WithError(err).Warn("error unlocking")
 116                                 }
 117                         }
 118                 }
 119                 // Shut down idle workers that didn't get any
 120                 // containers mapped onto them before we hit quota.
 121                 for it, n := range unalloc {
 122                         if n < 1 {
 123                                 continue
 124                         }
 125                         sch.pool.Shutdown(it)
 126                 }
 127         }
 128 }
 129
 130 // Lock the given container. Should be called in a new goroutine.
 131 func (sch *Scheduler) lockContainer(logger logrus.FieldLogger, uuid string) {
 132         if !sch.uuidLock(uuid, "lock") {
 133                 return
 134         }
 135         defer sch.uuidUnlock(uuid)
 136         if ctr, ok := sch.queue.Get(uuid); !ok || ctr.State != arvados.ContainerStateQueued {
 137                 // This happens if the container has been cancelled or
 138                 // locked since runQueue called sch.queue.Entries(),
 139                 // possibly by a lockContainer() call from a previous
 140                 // runQueue iteration. In any case, we will respond
 141                 // appropriately on the next runQueue iteration, which
 142                 // will have already been triggered by the queue
 143                 // update.
 144                 logger.WithField("State", ctr.State).Debug("container no longer queued by the time we decided to lock it, doing nothing")
 145                 return
 146         }
 147         err := sch.queue.Lock(uuid)
 148         if err != nil {
 149                 logger.WithError(err).Warn("error locking container")
 150                 return
 151         }
 152         logger.Debug("lock succeeded")
 153         ctr, ok := sch.queue.Get(uuid)
 154         if !ok {
 155                 logger.Error("(BUG?) container disappeared from queue after Lock succeeded")
 156         } else if ctr.State != arvados.ContainerStateLocked {
 157                 logger.Warnf("(race?) container has state=%q after Lock succeeded", ctr.State)
 158         }
 159 }
 160
 161 // Acquire a non-blocking lock for specified UUID, returning true if
 162 // successful.  The op argument is used only for debug logs.
 163 //
 164 // If the lock is not available, uuidLock arranges to wake up the
 165 // scheduler after a short delay, so it can retry whatever operation
 166 // is trying to get the lock (if that operation is still worth doing).
 167 //
 168 // This mechanism helps avoid spamming the controller/database with
 169 // concurrent updates for any single container, even when the
 170 // scheduler loop is running frequently.
 171 func (sch *Scheduler) uuidLock(uuid, op string) bool {
 172         sch.mtx.Lock()
 173         defer sch.mtx.Unlock()
 174         logger := sch.logger.WithFields(logrus.Fields{
 175                 "ContainerUUID": uuid,
 176                 "Op":            op,
 177         })
 178         if op, locked := sch.uuidOp[uuid]; locked {
 179                 logger.Debugf("uuidLock not available, Op=%s in progress", op)
 180                 // Make sure the scheduler loop wakes up to retry.
 181                 sch.wakeup.Reset(time.Second / 4)
 182                 return false
 183         }
 184         logger.Debug("uuidLock acquired")
 185         sch.uuidOp[uuid] = op
 186         return true
 187 }
 188
 189 func (sch *Scheduler) uuidUnlock(uuid string) {
 190         sch.mtx.Lock()
 191         defer sch.mtx.Unlock()
 192         delete(sch.uuidOp, uuid)
 193 }