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