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.curoverse.com/arvados.git/lib/dispatchcloud/container"
  12         "git.curoverse.com/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
  37 tryrun:
  38         for i, ctr := range sorted {
  39                 ctr, it := ctr.Container, ctr.InstanceType
  40                 logger := sch.logger.WithFields(logrus.Fields{
  41                         "ContainerUUID": ctr.UUID,
  42                         "InstanceType":  it.Name,
  43                 })
  44                 if _, running := running[ctr.UUID]; running || ctr.Priority < 1 {
  45                         continue
  46                 }
  47                 switch ctr.State {
  48                 case arvados.ContainerStateQueued:
  49                         if unalloc[it] < 1 && sch.pool.AtQuota() {
  50                                 logger.Debug("not locking: AtQuota and no unalloc workers")
  51                                 overquota = sorted[i:]
  52                                 break tryrun
  53                         }
  54                         go sch.lockContainer(logger, ctr.UUID)
  55                         unalloc[it]--
  56                 case arvados.ContainerStateLocked:
  57                         if unalloc[it] > 0 {
  58                                 unalloc[it]--
  59                         } else if sch.pool.AtQuota() {
  60                                 logger.Debug("not starting: AtQuota and no unalloc workers")
  61                                 overquota = sorted[i:]
  62                                 break tryrun
  63                         } else {
  64                                 logger.Info("creating new instance")
  65                                 if !sch.pool.Create(it) {
  66                                         // (Note pool.Create works
  67                                         // asynchronously and logs its
  68                                         // own failures, so we don't
  69                                         // need to log this as a
  70                                         // failure.)
  71
  72                                         sch.queue.Unlock(ctr.UUID)
  73                                         // Don't let lower-priority
  74                                         // containers starve this one
  75                                         // by using keeping idle
  76                                         // workers alive on different
  77                                         // instance types.  TODO:
  78                                         // avoid getting starved here
  79                                         // if instances of a specific
  80                                         // type always fail.
  81                                         overquota = sorted[i:]
  82                                         break tryrun
  83                                 }
  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.StartContainer(it, ctr) {
  92                                 // Success.
  93                         } else {
  94                                 dontstart[it] = true
  95                         }
  96                 }
  97         }
  98
  99         if len(overquota) > 0 {
 100                 // Unlock any containers that are unmappable while
 101                 // we're at quota.
 102                 for _, ctr := range overquota {
 103                         ctr := ctr.Container
 104                         if ctr.State == arvados.ContainerStateLocked {
 105                                 logger := sch.logger.WithField("ContainerUUID", ctr.UUID)
 106                                 logger.Debug("unlock because pool capacity is used by higher priority containers")
 107                                 err := sch.queue.Unlock(ctr.UUID)
 108                                 if err != nil {
 109                                         logger.WithError(err).Warn("error unlocking")
 110                                 }
 111                         }
 112                 }
 113                 // Shut down idle workers that didn't get any
 114                 // containers mapped onto them before we hit quota.
 115                 for it, n := range unalloc {
 116                         if n < 1 {
 117                                 continue
 118                         }
 119                         sch.pool.Shutdown(it)
 120                 }
 121         }
 122 }
 123
 124 // Lock the given container. Should be called in a new goroutine.
 125 func (sch *Scheduler) lockContainer(logger logrus.FieldLogger, uuid string) {
 126         if !sch.uuidLock(uuid, "lock") {
 127                 return
 128         }
 129         defer sch.uuidUnlock(uuid)
 130         if ctr, ok := sch.queue.Get(uuid); !ok || ctr.State != arvados.ContainerStateQueued {
 131                 // This happens if the container has been cancelled or
 132                 // locked since runQueue called sch.queue.Entries(),
 133                 // possibly by a lockContainer() call from a previous
 134                 // runQueue iteration. In any case, we will respond
 135                 // appropriately on the next runQueue iteration, which
 136                 // will have already been triggered by the queue
 137                 // update.
 138                 logger.WithField("State", ctr.State).Debug("container no longer queued by the time we decided to lock it, doing nothing")
 139                 return
 140         }
 141         err := sch.queue.Lock(uuid)
 142         if err != nil {
 143                 logger.WithError(err).Warn("error locking container")
 144                 return
 145         }
 146         logger.Debug("lock succeeded")
 147         ctr, ok := sch.queue.Get(uuid)
 148         if !ok {
 149                 logger.Error("(BUG?) container disappeared from queue after Lock succeeded")
 150         } else if ctr.State != arvados.ContainerStateLocked {
 151                 logger.Warnf("(race?) container has state=%q after Lock succeeded", ctr.State)
 152         }
 153 }
 154
 155 // Acquire a non-blocking lock for specified UUID, returning true if
 156 // successful.  The op argument is used only for debug logs.
 157 //
 158 // If the lock is not available, uuidLock arranges to wake up the
 159 // scheduler after a short delay, so it can retry whatever operation
 160 // is trying to get the lock (if that operation is still worth doing).
 161 //
 162 // This mechanism helps avoid spamming the controller/database with
 163 // concurrent updates for any single container, even when the
 164 // scheduler loop is running frequently.
 165 func (sch *Scheduler) uuidLock(uuid, op string) bool {
 166         sch.mtx.Lock()
 167         defer sch.mtx.Unlock()
 168         logger := sch.logger.WithFields(logrus.Fields{
 169                 "ContainerUUID": uuid,
 170                 "Op":            op,
 171         })
 172         if op, locked := sch.uuidOp[uuid]; locked {
 173                 logger.Debugf("uuidLock not available, Op=%s in progress", op)
 174                 // Make sure the scheduler loop wakes up to retry.
 175                 sch.wakeup.Reset(time.Second / 4)
 176                 return false
 177         }
 178         logger.Debug("uuidLock acquired")
 179         sch.uuidOp[uuid] = op
 180         return true
 181 }
 182
 183 func (sch *Scheduler) uuidUnlock(uuid string) {
 184         sch.mtx.Lock()
 185         defer sch.mtx.Unlock()
 186         delete(sch.uuidOp, uuid)
 187 }