1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
11 "git.arvados.org/arvados.git/lib/dispatchcloud/container"
12 "git.arvados.org/arvados.git/sdk/go/arvados"
13 "github.com/sirupsen/logrus"
16 var quietAfter503 = time.Minute
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)
24 sort.Slice(sorted, func(i, j int) bool {
25 if pi, pj := sorted[i].Container.Priority, sorted[j].Container.Priority; pi != pj {
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)
37 running := sch.pool.Running()
38 unalloc := sch.pool.Unallocated()
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.
46 if newlimit := len(running) - 1; newlimit < 1 {
47 sch.maxConcurrency = 1
49 sch.maxConcurrency = newlimit
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.
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
65 if sch.last503time.IsZero() {
66 sch.mLast503Time.Set(0)
68 sch.mLast503Time.Set(float64(sch.last503time.Unix()))
70 sch.mMaxContainerConcurrency.Set(float64(sch.maxConcurrency))
72 sch.logger.WithFields(logrus.Fields{
73 "Containers": len(sorted),
74 "Processes": len(running),
75 "maxConcurrency": sch.maxConcurrency,
78 dontstart := map[arvados.InstanceType]bool{}
79 var overquota []container.QueueEnt // entries that are unmappable because of worker pool quota
80 var containerAllocatedWorkerBootingCount int
82 // trying is #containers running + #containers we're trying to
83 // start. We stop trying to start more containers if this
84 // reaches the dynamic maxConcurrency limit.
85 trying := len(running)
90 for i, ctr := range sorted {
91 ctr, it := ctr.Container, ctr.InstanceType
92 logger := sch.logger.WithFields(logrus.Fields{
93 "ContainerUUID": ctr.UUID,
94 "InstanceType": it.Name,
96 if ctr.SchedulingParameters.Supervisor {
98 if sch.maxSupervisors > 0 && supervisors > sch.maxSupervisors {
102 if _, running := running[ctr.UUID]; running || ctr.Priority < 1 {
106 case arvados.ContainerStateQueued:
107 if sch.maxConcurrency > 0 && trying >= sch.maxConcurrency {
108 logger.Tracef("not locking: already at maxConcurrency %d", sch.maxConcurrency)
109 overquota = sorted[i:]
113 if unalloc[it] < 1 && sch.pool.AtQuota() {
114 logger.Trace("not locking: AtQuota and no unalloc workers")
115 overquota = sorted[i:]
118 if sch.pool.KillContainer(ctr.UUID, "about to lock") {
119 logger.Info("not locking: crunch-run process from previous attempt has not exited")
122 go sch.lockContainer(logger, ctr.UUID)
124 case arvados.ContainerStateLocked:
125 if sch.maxConcurrency > 0 && trying >= sch.maxConcurrency {
126 logger.Debugf("not starting: already at maxConcurrency %d", sch.maxConcurrency)
127 overquota = sorted[i:]
133 } else if sch.pool.AtQuota() {
134 // Don't let lower-priority containers
135 // starve this one by using keeping
136 // idle workers alive on different
138 logger.Trace("overquota")
139 overquota = sorted[i:]
141 } else if sch.pool.Create(it) {
142 // Success. (Note pool.Create works
143 // asynchronously and does its own
144 // logging about the eventual outcome,
145 // so we don't need to.)
146 logger.Info("creating new instance")
148 // Failed despite not being at quota,
149 // e.g., cloud ops throttled. TODO:
150 // avoid getting starved here if
151 // instances of a specific type always
153 logger.Trace("pool declined to create new instance")
158 // We already tried & failed to start
159 // a higher-priority container on the
160 // same instance type. Don't let this
161 // one sneak in ahead of it.
162 } else if sch.pool.KillContainer(ctr.UUID, "about to start") {
163 logger.Info("not restarting yet: crunch-run process from previous attempt has not exited")
164 } else if sch.pool.StartContainer(it, ctr) {
165 logger.Trace("StartContainer => true")
168 logger.Trace("StartContainer => false")
169 containerAllocatedWorkerBootingCount += 1
175 sch.mContainersAllocatedNotStarted.Set(float64(containerAllocatedWorkerBootingCount))
176 sch.mContainersNotAllocatedOverQuota.Set(float64(len(overquota)))
178 if len(overquota) > 0 {
179 // Unlock any containers that are unmappable while
180 // we're at quota (but if they have already been
181 // scheduled and they're loading docker images etc.,
183 for _, ctr := range overquota {
185 _, toolate := running[ctr.UUID]
186 if ctr.State == arvados.ContainerStateLocked && !toolate {
187 logger := sch.logger.WithField("ContainerUUID", ctr.UUID)
188 logger.Debug("unlock because pool capacity is used by higher priority containers")
189 err := sch.queue.Unlock(ctr.UUID)
191 logger.WithError(err).Warn("error unlocking")
195 // Shut down idle workers that didn't get any
196 // containers mapped onto them before we hit quota.
197 for it, n := range unalloc {
201 sch.pool.Shutdown(it)
206 // Lock the given container. Should be called in a new goroutine.
207 func (sch *Scheduler) lockContainer(logger logrus.FieldLogger, uuid string) {
208 if !sch.uuidLock(uuid, "lock") {
211 defer sch.uuidUnlock(uuid)
212 if ctr, ok := sch.queue.Get(uuid); !ok || ctr.State != arvados.ContainerStateQueued {
213 // This happens if the container has been cancelled or
214 // locked since runQueue called sch.queue.Entries(),
215 // possibly by a lockContainer() call from a previous
216 // runQueue iteration. In any case, we will respond
217 // appropriately on the next runQueue iteration, which
218 // will have already been triggered by the queue
220 logger.WithField("State", ctr.State).Debug("container no longer queued by the time we decided to lock it, doing nothing")
223 err := sch.queue.Lock(uuid)
225 logger.WithError(err).Warn("error locking container")
228 logger.Debug("lock succeeded")
229 ctr, ok := sch.queue.Get(uuid)
231 logger.Error("(BUG?) container disappeared from queue after Lock succeeded")
232 } else if ctr.State != arvados.ContainerStateLocked {
233 logger.Warnf("(race?) container has state=%q after Lock succeeded", ctr.State)
237 // Acquire a non-blocking lock for specified UUID, returning true if
238 // successful. The op argument is used only for debug logs.
240 // If the lock is not available, uuidLock arranges to wake up the
241 // scheduler after a short delay, so it can retry whatever operation
242 // is trying to get the lock (if that operation is still worth doing).
244 // This mechanism helps avoid spamming the controller/database with
245 // concurrent updates for any single container, even when the
246 // scheduler loop is running frequently.
247 func (sch *Scheduler) uuidLock(uuid, op string) bool {
249 defer sch.mtx.Unlock()
250 logger := sch.logger.WithFields(logrus.Fields{
251 "ContainerUUID": uuid,
254 if op, locked := sch.uuidOp[uuid]; locked {
255 logger.Debugf("uuidLock not available, Op=%s in progress", op)
256 // Make sure the scheduler loop wakes up to retry.
257 sch.wakeup.Reset(time.Second / 4)
260 logger.Debug("uuidLock acquired")
261 sch.uuidOp[uuid] = op
265 func (sch *Scheduler) uuidUnlock(uuid string) {
267 defer sch.mtx.Unlock()
268 delete(sch.uuidOp, uuid)