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 var quietAfter503 = time.Minute
  17
  18 func isSupervisor(ctr arvados.Container) bool {
  19         return (len(ctr.Command) > 0 && ctr.Command[0] == "arvados-cwl-runner") || ctr.SchedulingParameters.Supervisor
  20 }
  21
  22 func (sch *Scheduler) runQueue() {
  23         unsorted, _ := sch.queue.Entries()
  24         sorted := make([]container.QueueEnt, 0, len(unsorted))
  25         for _, ent := range unsorted {
  26                 sorted = append(sorted, ent)
  27         }
  28         sort.Slice(sorted, func(i, j int) bool {
  29                 if pi, pj := sorted[i].Container.Priority, sorted[j].Container.Priority; pi != pj {
  30                         return pi > pj
  31                 } else {
  32                         // When containers have identical priority,
  33                         // start them in the order we first noticed
  34                         // them. This avoids extra lock/unlock cycles
  35                         // when we unlock the containers that don't
  36                         // fit in the available pool.
  37                         return sorted[i].FirstSeenAt.Before(sorted[j].FirstSeenAt)
  38                 }
  39         })
  40
  41         running := sch.pool.Running()
  42         unalloc := sch.pool.Unallocated()
  43
  44         if t := sch.client.Last503(); t.After(sch.last503time) {
  45                 // API has sent an HTTP 503 response since last time
  46                 // we checked. Use current #containers - 1 as
  47                 // maxConcurrency, i.e., try to stay just below the
  48                 // level where we see 503s.
  49                 sch.last503time = t
  50                 if newlimit := len(running) - 1; newlimit < 1 {
  51                         sch.maxConcurrency = 1
  52                 } else {
  53                         sch.maxConcurrency = newlimit
  54                 }
  55         } else if sch.maxConcurrency > 0 && time.Since(sch.last503time) > quietAfter503 {
  56                 // If we haven't seen any 503 errors lately, raise
  57                 // limit to ~10% beyond the current workload.
  58                 //
  59                 // As we use the added 10% to schedule more
  60                 // containers, len(running) will increase and we'll
  61                 // push the limit up further. Soon enough,
  62                 // maxConcurrency will get high enough to schedule the
  63                 // entire queue, hit pool quota, or get 503s again.
  64                 max := len(running)*11/10 + 1
  65                 if sch.maxConcurrency < max {
  66                         sch.maxConcurrency = max
  67                 }
  68         }
  69         if sch.last503time.IsZero() {
  70                 sch.mLast503Time.Set(0)
  71         } else {
  72                 sch.mLast503Time.Set(float64(sch.last503time.Unix()))
  73         }
  74         sch.mMaxContainerConcurrency.Set(float64(sch.maxConcurrency))
  75
  76         sch.logger.WithFields(logrus.Fields{
  77                 "Containers":     len(sorted),
  78                 "Processes":      len(running),
  79                 "maxConcurrency": sch.maxConcurrency,
  80         }).Debug("runQueue")
  81
  82         dontstart := map[arvados.InstanceType]bool{}
  83         var overquota []container.QueueEnt // entries that are unmappable because of worker pool quota
  84         var containerAllocatedWorkerBootingCount int
  85
  86         // trying is #containers running + #containers we're trying to
  87         // start. We stop trying to start more containers if this
  88         // reaches the dynamic maxConcurrency limit.
  89         trying := len(running)
  90
  91         supervisors := 0
  92
  93 tryrun:
  94         for i, ctr := range sorted {
  95                 ctr, it := ctr.Container, ctr.InstanceType
  96                 logger := sch.logger.WithFields(logrus.Fields{
  97                         "ContainerUUID": ctr.UUID,
  98                         "InstanceType":  it.Name,
  99                 })
 100                 if isSupervisor(ctr) {
 101                         supervisors += 1
 102                         if sch.maxSupervisors > 0 && supervisors > sch.maxSupervisors {
 103                                 continue
 104                         }
 105                 }
 106                 if _, running := running[ctr.UUID]; running || ctr.Priority < 1 {
 107                         continue
 108                 }
 109                 switch ctr.State {
 110                 case arvados.ContainerStateQueued:
 111                         if sch.maxConcurrency > 0 && trying >= sch.maxConcurrency {
 112                                 logger.Tracef("not locking: already at maxConcurrency %d", sch.maxConcurrency)
 113                                 overquota = sorted[i:]
 114                                 break tryrun
 115                         }
 116                         trying++
 117                         if unalloc[it] < 1 && sch.pool.AtQuota() {
 118                                 logger.Trace("not locking: AtQuota and no unalloc workers")
 119                                 overquota = sorted[i:]
 120                                 break tryrun
 121                         }
 122                         if sch.pool.KillContainer(ctr.UUID, "about to lock") {
 123                                 logger.Info("not locking: crunch-run process from previous attempt has not exited")
 124                                 continue
 125                         }
 126                         go sch.lockContainer(logger, ctr.UUID)
 127                         unalloc[it]--
 128                 case arvados.ContainerStateLocked:
 129                         if sch.maxConcurrency > 0 && trying >= sch.maxConcurrency {
 130                                 logger.Debugf("not starting: already at maxConcurrency %d", sch.maxConcurrency)
 131                                 overquota = sorted[i:]
 132                                 break tryrun
 133                         }
 134                         trying++
 135                         if unalloc[it] > 0 {
 136                                 unalloc[it]--
 137                         } else if sch.pool.AtQuota() {
 138                                 // Don't let lower-priority containers
 139                                 // starve this one by using keeping
 140                                 // idle workers alive on different
 141                                 // instance types.
 142                                 logger.Trace("overquota")
 143                                 overquota = sorted[i:]
 144                                 break tryrun
 145                         } else if sch.pool.Create(it) {
 146                                 // Success. (Note pool.Create works
 147                                 // asynchronously and does its own
 148                                 // logging about the eventual outcome,
 149                                 // so we don't need to.)
 150                                 logger.Info("creating new instance")
 151                         } else {
 152                                 // Failed despite not being at quota,
 153                                 // e.g., cloud ops throttled.  TODO:
 154                                 // avoid getting starved here if
 155                                 // instances of a specific type always
 156                                 // fail.
 157                                 logger.Trace("pool declined to create new instance")
 158                                 continue
 159                         }
 160
 161                         if dontstart[it] {
 162                                 // We already tried & failed to start
 163                                 // a higher-priority container on the
 164                                 // same instance type. Don't let this
 165                                 // one sneak in ahead of it.
 166                         } else if sch.pool.KillContainer(ctr.UUID, "about to start") {
 167                                 logger.Info("not restarting yet: crunch-run process from previous attempt has not exited")
 168                         } else if sch.pool.StartContainer(it, ctr) {
 169                                 logger.Trace("StartContainer => true")
 170                                 // Success.
 171                         } else {
 172                                 logger.Trace("StartContainer => false")
 173                                 containerAllocatedWorkerBootingCount += 1
 174                                 dontstart[it] = true
 175                         }
 176                 }
 177         }
 178
 179         sch.mContainersAllocatedNotStarted.Set(float64(containerAllocatedWorkerBootingCount))
 180         sch.mContainersNotAllocatedOverQuota.Set(float64(len(overquota)))
 181
 182         if len(overquota) > 0 {
 183                 // Unlock any containers that are unmappable while
 184                 // we're at quota (but if they have already been
 185                 // scheduled and they're loading docker images etc.,
 186                 // let them run).
 187                 for _, ctr := range overquota {
 188                         ctr := ctr.Container
 189                         _, toolate := running[ctr.UUID]
 190                         if ctr.State == arvados.ContainerStateLocked && !toolate {
 191                                 logger := sch.logger.WithField("ContainerUUID", ctr.UUID)
 192                                 logger.Debug("unlock because pool capacity is used by higher priority containers")
 193                                 err := sch.queue.Unlock(ctr.UUID)
 194                                 if err != nil {
 195                                         logger.WithError(err).Warn("error unlocking")
 196                                 }
 197                         }
 198                 }
 199                 // Shut down idle workers that didn't get any
 200                 // containers mapped onto them before we hit quota.
 201                 for it, n := range unalloc {
 202                         if n < 1 {
 203                                 continue
 204                         }
 205                         sch.pool.Shutdown(it)
 206                 }
 207         }
 208 }
 209
 210 // Lock the given container. Should be called in a new goroutine.
 211 func (sch *Scheduler) lockContainer(logger logrus.FieldLogger, uuid string) {
 212         if !sch.uuidLock(uuid, "lock") {
 213                 return
 214         }
 215         defer sch.uuidUnlock(uuid)
 216         if ctr, ok := sch.queue.Get(uuid); !ok || ctr.State != arvados.ContainerStateQueued {
 217                 // This happens if the container has been cancelled or
 218                 // locked since runQueue called sch.queue.Entries(),
 219                 // possibly by a lockContainer() call from a previous
 220                 // runQueue iteration. In any case, we will respond
 221                 // appropriately on the next runQueue iteration, which
 222                 // will have already been triggered by the queue
 223                 // update.
 224                 logger.WithField("State", ctr.State).Debug("container no longer queued by the time we decided to lock it, doing nothing")
 225                 return
 226         }
 227         err := sch.queue.Lock(uuid)
 228         if err != nil {
 229                 logger.WithError(err).Warn("error locking container")
 230                 return
 231         }
 232         logger.Debug("lock succeeded")
 233         ctr, ok := sch.queue.Get(uuid)
 234         if !ok {
 235                 logger.Error("(BUG?) container disappeared from queue after Lock succeeded")
 236         } else if ctr.State != arvados.ContainerStateLocked {
 237                 logger.Warnf("(race?) container has state=%q after Lock succeeded", ctr.State)
 238         }
 239 }
 240
 241 // Acquire a non-blocking lock for specified UUID, returning true if
 242 // successful.  The op argument is used only for debug logs.
 243 //
 244 // If the lock is not available, uuidLock arranges to wake up the
 245 // scheduler after a short delay, so it can retry whatever operation
 246 // is trying to get the lock (if that operation is still worth doing).
 247 //
 248 // This mechanism helps avoid spamming the controller/database with
 249 // concurrent updates for any single container, even when the
 250 // scheduler loop is running frequently.
 251 func (sch *Scheduler) uuidLock(uuid, op string) bool {
 252         sch.mtx.Lock()
 253         defer sch.mtx.Unlock()
 254         logger := sch.logger.WithFields(logrus.Fields{
 255                 "ContainerUUID": uuid,
 256                 "Op":            op,
 257         })
 258         if op, locked := sch.uuidOp[uuid]; locked {
 259                 logger.Debugf("uuidLock not available, Op=%s in progress", op)
 260                 // Make sure the scheduler loop wakes up to retry.
 261                 sch.wakeup.Reset(time.Second / 4)
 262                 return false
 263         }
 264         logger.Debug("uuidLock acquired")
 265         sch.uuidOp[uuid] = op
 266         return true
 267 }
 268
 269 func (sch *Scheduler) uuidUnlock(uuid string) {
 270         sch.mtx.Lock()
 271         defer sch.mtx.Unlock()
 272         delete(sch.uuidOp, uuid)
 273 }