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 (sch *Scheduler) runQueue() {
  19         running := sch.pool.Running()
  20         unalloc := sch.pool.Unallocated()
  21
  22         unsorted, _ := sch.queue.Entries()
  23         sorted := make([]container.QueueEnt, 0, len(unsorted))
  24         for _, ent := range unsorted {
  25                 sorted = append(sorted, ent)
  26         }
  27         sort.Slice(sorted, func(i, j int) bool {
  28                 _, irunning := running[sorted[i].Container.UUID]
  29                 _, jrunning := running[sorted[j].Container.UUID]
  30                 if irunning != jrunning {
  31                         // Ensure the "tryrun" loop (see below) sees
  32                         // already-scheduled containers first, to
  33                         // ensure existing supervisor containers are
  34                         // properly counted before we decide whether
  35                         // we have room for new ones.
  36                         return irunning
  37                 }
  38                 ilocked := sorted[i].Container.State == arvados.ContainerStateLocked
  39                 jlocked := sorted[j].Container.State == arvados.ContainerStateLocked
  40                 if ilocked != jlocked {
  41                         // Give precedence to containers that we have
  42                         // already locked, even if higher-priority
  43                         // containers have since arrived in the
  44                         // queue. This avoids undesirable queue churn
  45                         // effects including extra lock/unlock cycles
  46                         // and bringing up new instances and quickly
  47                         // shutting them down to make room for
  48                         // different instance sizes.
  49                         return ilocked
  50                 } else if pi, pj := sorted[i].Container.Priority, sorted[j].Container.Priority; pi != pj {
  51                         return pi > pj
  52                 } else {
  53                         // When containers have identical priority,
  54                         // start them in the order we first noticed
  55                         // them. This avoids extra lock/unlock cycles
  56                         // when we unlock the containers that don't
  57                         // fit in the available pool.
  58                         return sorted[i].FirstSeenAt.Before(sorted[j].FirstSeenAt)
  59                 }
  60         })
  61
  62         if t := sch.client.Last503(); t.After(sch.last503time) {
  63                 // API has sent an HTTP 503 response since last time
  64                 // we checked. Use current #containers - 1 as
  65                 // maxConcurrency, i.e., try to stay just below the
  66                 // level where we see 503s.
  67                 sch.last503time = t
  68                 if newlimit := len(running) - 1; newlimit < 1 {
  69                         sch.maxConcurrency = 1
  70                 } else {
  71                         sch.maxConcurrency = newlimit
  72                 }
  73         } else if sch.maxConcurrency > 0 && time.Since(sch.last503time) > quietAfter503 {
  74                 // If we haven't seen any 503 errors lately, raise
  75                 // limit to ~10% beyond the current workload.
  76                 //
  77                 // As we use the added 10% to schedule more
  78                 // containers, len(running) will increase and we'll
  79                 // push the limit up further. Soon enough,
  80                 // maxConcurrency will get high enough to schedule the
  81                 // entire queue, hit pool quota, or get 503s again.
  82                 max := len(running)*11/10 + 1
  83                 if sch.maxConcurrency < max {
  84                         sch.maxConcurrency = max
  85                 }
  86         }
  87         if sch.last503time.IsZero() {
  88                 sch.mLast503Time.Set(0)
  89         } else {
  90                 sch.mLast503Time.Set(float64(sch.last503time.Unix()))
  91         }
  92         sch.mMaxContainerConcurrency.Set(float64(sch.maxConcurrency))
  93
  94         sch.logger.WithFields(logrus.Fields{
  95                 "Containers":     len(sorted),
  96                 "Processes":      len(running),
  97                 "maxConcurrency": sch.maxConcurrency,
  98         }).Debug("runQueue")
  99
 100         dontstart := map[arvados.InstanceType]bool{}
 101         var overquota []container.QueueEnt    // entries that are unmappable because of worker pool quota
 102         var overmaxsuper []container.QueueEnt // unmappable because max supervisors (these are not included in overquota)
 103         var containerAllocatedWorkerBootingCount int
 104
 105         // trying is #containers running + #containers we're trying to
 106         // start. We stop trying to start more containers if this
 107         // reaches the dynamic maxConcurrency limit.
 108         trying := len(running)
 109
 110         supervisors := 0
 111
 112 tryrun:
 113         for i, ent := range sorted {
 114                 ctr, it := ent.Container, ent.InstanceType
 115                 logger := sch.logger.WithFields(logrus.Fields{
 116                         "ContainerUUID": ctr.UUID,
 117                         "InstanceType":  it.Name,
 118                 })
 119                 if ctr.SchedulingParameters.Supervisor {
 120                         supervisors += 1
 121                         if sch.maxSupervisors > 0 && supervisors > sch.maxSupervisors {
 122                                 overmaxsuper = append(overmaxsuper, sorted[i])
 123                                 continue
 124                         }
 125                 }
 126                 if _, running := running[ctr.UUID]; running || ctr.Priority < 1 {
 127                         continue
 128                 }
 129                 switch ctr.State {
 130                 case arvados.ContainerStateQueued:
 131                         if sch.maxConcurrency > 0 && trying >= sch.maxConcurrency {
 132                                 logger.Tracef("not locking: already at maxConcurrency %d", sch.maxConcurrency)
 133                                 continue
 134                         }
 135                         trying++
 136                         if unalloc[it] < 1 && sch.pool.AtQuota() {
 137                                 logger.Trace("not locking: AtQuota and no unalloc workers")
 138                                 overquota = sorted[i:]
 139                                 break tryrun
 140                         }
 141                         if sch.pool.KillContainer(ctr.UUID, "about to lock") {
 142                                 logger.Info("not locking: crunch-run process from previous attempt has not exited")
 143                                 continue
 144                         }
 145                         go sch.lockContainer(logger, ctr.UUID)
 146                         unalloc[it]--
 147                 case arvados.ContainerStateLocked:
 148                         if sch.maxConcurrency > 0 && trying >= sch.maxConcurrency {
 149                                 logger.Tracef("not starting: already at maxConcurrency %d", sch.maxConcurrency)
 150                                 continue
 151                         }
 152                         trying++
 153                         if unalloc[it] > 0 {
 154                                 unalloc[it]--
 155                         } else if sch.pool.AtQuota() {
 156                                 // Don't let lower-priority containers
 157                                 // starve this one by using keeping
 158                                 // idle workers alive on different
 159                                 // instance types.
 160                                 logger.Trace("overquota")
 161                                 overquota = sorted[i:]
 162                                 break tryrun
 163                         } else if sch.pool.Create(it) {
 164                                 // Success. (Note pool.Create works
 165                                 // asynchronously and does its own
 166                                 // logging about the eventual outcome,
 167                                 // so we don't need to.)
 168                                 logger.Info("creating new instance")
 169                         } else {
 170                                 // Failed despite not being at quota,
 171                                 // e.g., cloud ops throttled.  TODO:
 172                                 // avoid getting starved here if
 173                                 // instances of a specific type always
 174                                 // fail.
 175                                 logger.Trace("pool declined to create new instance")
 176                                 continue
 177                         }
 178
 179                         if dontstart[it] {
 180                                 // We already tried & failed to start
 181                                 // a higher-priority container on the
 182                                 // same instance type. Don't let this
 183                                 // one sneak in ahead of it.
 184                         } else if sch.pool.KillContainer(ctr.UUID, "about to start") {
 185                                 logger.Info("not restarting yet: crunch-run process from previous attempt has not exited")
 186                         } else if sch.pool.StartContainer(it, ctr) {
 187                                 logger.Trace("StartContainer => true")
 188                                 // Success.
 189                         } else {
 190                                 logger.Trace("StartContainer => false")
 191                                 containerAllocatedWorkerBootingCount += 1
 192                                 dontstart[it] = true
 193                         }
 194                 }
 195         }
 196
 197         sch.mContainersAllocatedNotStarted.Set(float64(containerAllocatedWorkerBootingCount))
 198         sch.mContainersNotAllocatedOverQuota.Set(float64(len(overquota) + len(overmaxsuper)))
 199
 200         if len(overquota)+len(overmaxsuper) > 0 {
 201                 // Unlock any containers that are unmappable while
 202                 // we're at quota (but if they have already been
 203                 // scheduled and they're loading docker images etc.,
 204                 // let them run).
 205                 for _, ctr := range append(overmaxsuper, overquota...) {
 206                         ctr := ctr.Container
 207                         _, toolate := running[ctr.UUID]
 208                         if ctr.State == arvados.ContainerStateLocked && !toolate {
 209                                 logger := sch.logger.WithField("ContainerUUID", ctr.UUID)
 210                                 logger.Info("unlock because pool capacity is used by higher priority containers")
 211                                 err := sch.queue.Unlock(ctr.UUID)
 212                                 if err != nil {
 213                                         logger.WithError(err).Warn("error unlocking")
 214                                 }
 215                         }
 216                 }
 217         }
 218         if len(overquota) > 0 {
 219                 // Shut down idle workers that didn't get any
 220                 // containers mapped onto them before we hit quota.
 221                 for it, n := range unalloc {
 222                         if n < 1 {
 223                                 continue
 224                         }
 225                         sch.pool.Shutdown(it)
 226                 }
 227         }
 228 }
 229
 230 // Lock the given container. Should be called in a new goroutine.
 231 func (sch *Scheduler) lockContainer(logger logrus.FieldLogger, uuid string) {
 232         if !sch.uuidLock(uuid, "lock") {
 233                 return
 234         }
 235         defer sch.uuidUnlock(uuid)
 236         if ctr, ok := sch.queue.Get(uuid); !ok || ctr.State != arvados.ContainerStateQueued {
 237                 // This happens if the container has been cancelled or
 238                 // locked since runQueue called sch.queue.Entries(),
 239                 // possibly by a lockContainer() call from a previous
 240                 // runQueue iteration. In any case, we will respond
 241                 // appropriately on the next runQueue iteration, which
 242                 // will have already been triggered by the queue
 243                 // update.
 244                 logger.WithField("State", ctr.State).Debug("container no longer queued by the time we decided to lock it, doing nothing")
 245                 return
 246         }
 247         err := sch.queue.Lock(uuid)
 248         if err != nil {
 249                 logger.WithError(err).Warn("error locking container")
 250                 return
 251         }
 252         logger.Debug("lock succeeded")
 253         ctr, ok := sch.queue.Get(uuid)
 254         if !ok {
 255                 logger.Error("(BUG?) container disappeared from queue after Lock succeeded")
 256         } else if ctr.State != arvados.ContainerStateLocked {
 257                 logger.Warnf("(race?) container has state=%q after Lock succeeded", ctr.State)
 258         }
 259 }
 260
 261 // Acquire a non-blocking lock for specified UUID, returning true if
 262 // successful.  The op argument is used only for debug logs.
 263 //
 264 // If the lock is not available, uuidLock arranges to wake up the
 265 // scheduler after a short delay, so it can retry whatever operation
 266 // is trying to get the lock (if that operation is still worth doing).
 267 //
 268 // This mechanism helps avoid spamming the controller/database with
 269 // concurrent updates for any single container, even when the
 270 // scheduler loop is running frequently.
 271 func (sch *Scheduler) uuidLock(uuid, op string) bool {
 272         sch.mtx.Lock()
 273         defer sch.mtx.Unlock()
 274         logger := sch.logger.WithFields(logrus.Fields{
 275                 "ContainerUUID": uuid,
 276                 "Op":            op,
 277         })
 278         if op, locked := sch.uuidOp[uuid]; locked {
 279                 logger.Debugf("uuidLock not available, Op=%s in progress", op)
 280                 // Make sure the scheduler loop wakes up to retry.
 281                 sch.wakeup.Reset(time.Second / 4)
 282                 return false
 283         }
 284         logger.Debug("uuidLock acquired")
 285         sch.uuidOp[uuid] = op
 286         return true
 287 }
 288
 289 func (sch *Scheduler) uuidUnlock(uuid string) {
 290         sch.mtx.Lock()
 291         defer sch.mtx.Unlock()
 292         delete(sch.uuidOp, uuid)
 293 }