forked from moby/swarmkit
/
scheduler.go
693 lines (601 loc) · 19.2 KB
/
scheduler.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
package scheduler
import (
"container/list"
"time"
"github.com/docker/swarmkit/api"
"github.com/docker/swarmkit/log"
"github.com/docker/swarmkit/manager/state"
"github.com/docker/swarmkit/manager/state/store"
"github.com/docker/swarmkit/protobuf/ptypes"
"golang.org/x/net/context"
)
const (
// monitorFailures is the lookback period for counting failures of
// a task to determine if a node is faulty for a particular service.
monitorFailures = 5 * time.Minute
// maxFailures is the number of failures within monitorFailures that
// triggers downweighting of a node in the sorting function.
maxFailures = 5
)
type schedulingDecision struct {
old *api.Task
new *api.Task
}
// Scheduler assigns tasks to nodes.
type Scheduler struct {
store *store.MemoryStore
unassignedTasks *list.List
// preassignedTasks already have NodeID, need resource validation
preassignedTasks map[string]*api.Task
nodeSet nodeSet
allTasks map[string]*api.Task
pipeline *Pipeline
// stopChan signals to the state machine to stop running
stopChan chan struct{}
// doneChan is closed when the state machine terminates
doneChan chan struct{}
}
// New creates a new scheduler.
func New(store *store.MemoryStore) *Scheduler {
return &Scheduler{
store: store,
unassignedTasks: list.New(),
preassignedTasks: make(map[string]*api.Task),
allTasks: make(map[string]*api.Task),
stopChan: make(chan struct{}),
doneChan: make(chan struct{}),
pipeline: NewPipeline(),
}
}
func (s *Scheduler) setupTasksList(tx store.ReadTx) error {
tasks, err := store.FindTasks(tx, store.All)
if err != nil {
return err
}
tasksByNode := make(map[string]map[string]*api.Task)
for _, t := range tasks {
// Ignore all tasks that have not reached PENDING
// state and tasks that no longer consume resources.
if t.Status.State < api.TaskStatePending || t.Status.State > api.TaskStateRunning {
continue
}
s.allTasks[t.ID] = t
if t.NodeID == "" {
s.enqueue(t)
continue
}
// preassigned tasks need to validate resource requirement on corresponding node
if t.Status.State == api.TaskStatePending {
s.preassignedTasks[t.ID] = t
continue
}
if tasksByNode[t.NodeID] == nil {
tasksByNode[t.NodeID] = make(map[string]*api.Task)
}
tasksByNode[t.NodeID][t.ID] = t
}
if err := s.buildNodeSet(tx, tasksByNode); err != nil {
return err
}
return nil
}
// Run is the scheduler event loop.
func (s *Scheduler) Run(ctx context.Context) error {
defer close(s.doneChan)
updates, cancel, err := store.ViewAndWatch(s.store, s.setupTasksList)
if err != nil {
log.G(ctx).WithError(err).Errorf("snapshot store update failed")
return err
}
defer cancel()
// Validate resource for tasks from preassigned tasks
// do this before other tasks because preassigned tasks like
// global service should start before other tasks
s.processPreassignedTasks(ctx)
// Queue all unassigned tasks before processing changes.
s.tick(ctx)
const (
// commitDebounceGap is the amount of time to wait between
// commit events to debounce them.
commitDebounceGap = 50 * time.Millisecond
// maxLatency is a time limit on the debouncing.
maxLatency = time.Second
)
var (
debouncingStarted time.Time
commitDebounceTimer *time.Timer
commitDebounceTimeout <-chan time.Time
)
pendingChanges := 0
schedule := func() {
if len(s.preassignedTasks) > 0 {
s.processPreassignedTasks(ctx)
}
if pendingChanges > 0 {
s.tick(ctx)
pendingChanges = 0
}
}
// Watch for changes.
for {
select {
case event := <-updates:
switch v := event.(type) {
case state.EventCreateTask:
pendingChanges += s.createTask(ctx, v.Task)
case state.EventUpdateTask:
pendingChanges += s.updateTask(ctx, v.Task)
case state.EventDeleteTask:
s.deleteTask(ctx, v.Task)
case state.EventCreateNode:
s.createOrUpdateNode(v.Node)
pendingChanges++
case state.EventUpdateNode:
s.createOrUpdateNode(v.Node)
pendingChanges++
case state.EventDeleteNode:
s.nodeSet.remove(v.Node.ID)
case state.EventCommit:
if commitDebounceTimer != nil {
if time.Since(debouncingStarted) > maxLatency {
commitDebounceTimer.Stop()
commitDebounceTimer = nil
commitDebounceTimeout = nil
schedule()
} else {
commitDebounceTimer.Reset(commitDebounceGap)
}
} else {
commitDebounceTimer = time.NewTimer(commitDebounceGap)
commitDebounceTimeout = commitDebounceTimer.C
debouncingStarted = time.Now()
}
}
case <-commitDebounceTimeout:
schedule()
commitDebounceTimer = nil
commitDebounceTimeout = nil
case <-s.stopChan:
return nil
}
}
}
// Stop causes the scheduler event loop to stop running.
func (s *Scheduler) Stop() {
close(s.stopChan)
<-s.doneChan
}
// enqueue queues a task for scheduling.
func (s *Scheduler) enqueue(t *api.Task) {
s.unassignedTasks.PushBack(t)
}
func (s *Scheduler) createTask(ctx context.Context, t *api.Task) int {
// Ignore all tasks that have not reached PENDING
// state, and tasks that no longer consume resources.
if t.Status.State < api.TaskStatePending || t.Status.State > api.TaskStateRunning {
return 0
}
s.allTasks[t.ID] = t
if t.NodeID == "" {
// unassigned task
s.enqueue(t)
return 1
}
if t.Status.State == api.TaskStatePending {
s.preassignedTasks[t.ID] = t
// preassigned tasks do not contribute to running tasks count
return 0
}
nodeInfo, err := s.nodeSet.nodeInfo(t.NodeID)
if err == nil && nodeInfo.addTask(t) {
s.nodeSet.updateNode(nodeInfo)
}
return 0
}
func (s *Scheduler) updateTask(ctx context.Context, t *api.Task) int {
// Ignore all tasks that have not reached PENDING
// state.
if t.Status.State < api.TaskStatePending {
return 0
}
oldTask := s.allTasks[t.ID]
// Ignore all tasks that have not reached ALLOCATED
// state, and tasks that no longer consume resources.
if t.Status.State > api.TaskStateRunning {
if oldTask == nil {
return 1
}
s.deleteTask(ctx, oldTask)
if t.Status.State != oldTask.Status.State &&
(t.Status.State == api.TaskStateFailed || t.Status.State == api.TaskStateRejected) {
nodeInfo, err := s.nodeSet.nodeInfo(t.NodeID)
if err == nil {
nodeInfo.taskFailed(ctx, t.ServiceID)
s.nodeSet.updateNode(nodeInfo)
}
}
return 1
}
if t.NodeID == "" {
// unassigned task
if oldTask != nil {
s.deleteTask(ctx, oldTask)
}
s.allTasks[t.ID] = t
s.enqueue(t)
return 1
}
if t.Status.State == api.TaskStatePending {
if oldTask != nil {
s.deleteTask(ctx, oldTask)
}
s.allTasks[t.ID] = t
s.preassignedTasks[t.ID] = t
// preassigned tasks do not contribute to running tasks count
return 0
}
s.allTasks[t.ID] = t
nodeInfo, err := s.nodeSet.nodeInfo(t.NodeID)
if err == nil && nodeInfo.addTask(t) {
s.nodeSet.updateNode(nodeInfo)
}
return 0
}
func (s *Scheduler) deleteTask(ctx context.Context, t *api.Task) {
delete(s.allTasks, t.ID)
delete(s.preassignedTasks, t.ID)
nodeInfo, err := s.nodeSet.nodeInfo(t.NodeID)
if err == nil && nodeInfo.removeTask(t) {
s.nodeSet.updateNode(nodeInfo)
}
}
func (s *Scheduler) createOrUpdateNode(n *api.Node) {
nodeInfo, _ := s.nodeSet.nodeInfo(n.ID)
var resources api.Resources
if n.Description != nil && n.Description.Resources != nil {
resources = *n.Description.Resources
// reconcile resources by looping over all tasks in this node
for _, task := range nodeInfo.Tasks {
reservations := taskReservations(task.Spec)
resources.MemoryBytes -= reservations.MemoryBytes
resources.NanoCPUs -= reservations.NanoCPUs
}
}
nodeInfo.Node = n
nodeInfo.AvailableResources = resources
s.nodeSet.addOrUpdateNode(nodeInfo)
}
func (s *Scheduler) processPreassignedTasks(ctx context.Context) {
schedulingDecisions := make(map[string]schedulingDecision, len(s.preassignedTasks))
for _, t := range s.preassignedTasks {
newT := s.taskFitNode(ctx, t, t.NodeID)
if newT == nil {
continue
}
schedulingDecisions[t.ID] = schedulingDecision{old: t, new: newT}
}
successful, failed := s.applySchedulingDecisions(ctx, schedulingDecisions)
for _, decision := range successful {
if decision.new.Status.State == api.TaskStateAssigned {
delete(s.preassignedTasks, decision.old.ID)
}
}
for _, decision := range failed {
s.allTasks[decision.old.ID] = decision.old
nodeInfo, err := s.nodeSet.nodeInfo(decision.new.NodeID)
if err == nil && nodeInfo.removeTask(decision.new) {
s.nodeSet.updateNode(nodeInfo)
}
}
}
// tick attempts to schedule the queue.
func (s *Scheduler) tick(ctx context.Context) {
tasksByCommonSpec := make(map[string]map[string]*api.Task)
schedulingDecisions := make(map[string]schedulingDecision, s.unassignedTasks.Len())
var next *list.Element
for e := s.unassignedTasks.Front(); e != nil; e = next {
next = e.Next()
t := s.allTasks[e.Value.(*api.Task).ID]
if t == nil || t.NodeID != "" {
// task deleted or already assigned
s.unassignedTasks.Remove(e)
continue
}
// Group common tasks with common specs by marshalling the spec
// into taskKey and using it as a map key.
// TODO(aaronl): Once specs are versioned, this will allow a
// much more efficient fast path.
fieldsToMarshal := api.Task{
ServiceID: t.ServiceID,
Spec: t.Spec,
}
marshalled, err := fieldsToMarshal.Marshal()
if err != nil {
panic(err)
}
taskGroupKey := string(marshalled)
if tasksByCommonSpec[taskGroupKey] == nil {
tasksByCommonSpec[taskGroupKey] = make(map[string]*api.Task)
}
tasksByCommonSpec[taskGroupKey][t.ID] = t
s.unassignedTasks.Remove(e)
}
for _, taskGroup := range tasksByCommonSpec {
s.scheduleTaskGroup(ctx, taskGroup, schedulingDecisions)
}
_, failed := s.applySchedulingDecisions(ctx, schedulingDecisions)
for _, decision := range failed {
s.allTasks[decision.old.ID] = decision.old
nodeInfo, err := s.nodeSet.nodeInfo(decision.new.NodeID)
if err == nil && nodeInfo.removeTask(decision.new) {
s.nodeSet.updateNode(nodeInfo)
}
// enqueue task for next scheduling attempt
s.enqueue(decision.old)
}
}
func (s *Scheduler) applySchedulingDecisions(ctx context.Context, schedulingDecisions map[string]schedulingDecision) (successful, failed []schedulingDecision) {
if len(schedulingDecisions) == 0 {
return
}
successful = make([]schedulingDecision, 0, len(schedulingDecisions))
// Apply changes to master store
applied, err := s.store.Batch(func(batch *store.Batch) error {
for len(schedulingDecisions) > 0 {
err := batch.Update(func(tx store.Tx) error {
// Update exactly one task inside this Update
// callback.
for taskID, decision := range schedulingDecisions {
delete(schedulingDecisions, taskID)
t := store.GetTask(tx, taskID)
if t == nil {
// Task no longer exists
nodeInfo, err := s.nodeSet.nodeInfo(decision.new.NodeID)
if err == nil && nodeInfo.removeTask(decision.new) {
s.nodeSet.updateNode(nodeInfo)
}
delete(s.allTasks, decision.old.ID)
continue
}
if t.Status.State == decision.new.Status.State && t.Status.Message == decision.new.Status.Message {
// No changes, ignore
continue
}
if t.Status.State >= api.TaskStateAssigned {
nodeInfo, err := s.nodeSet.nodeInfo(decision.new.NodeID)
if err != nil {
failed = append(failed, decision)
continue
}
node := store.GetNode(tx, decision.new.NodeID)
if node == nil || node.Meta.Version != nodeInfo.Meta.Version {
// node is out of date
failed = append(failed, decision)
continue
}
}
if err := store.UpdateTask(tx, decision.new); err != nil {
log.G(ctx).Debugf("scheduler failed to update task %s; will retry", taskID)
failed = append(failed, decision)
continue
}
successful = append(successful, decision)
return nil
}
return nil
})
if err != nil {
return err
}
}
return nil
})
if err != nil {
log.G(ctx).WithError(err).Error("scheduler tick transaction failed")
failed = append(failed, successful[applied:]...)
successful = successful[:applied]
}
return
}
// taskFitNode checks if a node has enough resources to accommodate a task.
func (s *Scheduler) taskFitNode(ctx context.Context, t *api.Task, nodeID string) *api.Task {
nodeInfo, err := s.nodeSet.nodeInfo(nodeID)
if err != nil {
// node does not exist in set (it may have been deleted)
return nil
}
newT := *t
s.pipeline.SetTask(t)
if !s.pipeline.Process(&nodeInfo) {
// this node cannot accommodate this task
newT.Status.Timestamp = ptypes.MustTimestampProto(time.Now())
newT.Status.Message = s.pipeline.Explain()
s.allTasks[t.ID] = &newT
return &newT
}
newT.Status = api.TaskStatus{
State: api.TaskStateAssigned,
Timestamp: ptypes.MustTimestampProto(time.Now()),
Message: "scheduler confirmed task can run on preassigned node",
}
s.allTasks[t.ID] = &newT
if nodeInfo.addTask(&newT) {
s.nodeSet.updateNode(nodeInfo)
}
return &newT
}
// scheduleTaskGroup schedules a batch of tasks that are part of the same
// service and share the same version of the spec.
func (s *Scheduler) scheduleTaskGroup(ctx context.Context, taskGroup map[string]*api.Task, schedulingDecisions map[string]schedulingDecision) {
// Pick at task at random from taskGroup to use for constraint
// evaluation. It doesn't matter which one we pick because all the
// tasks in the group are equal in terms of the fields the constraint
// filters consider.
var t *api.Task
for _, t = range taskGroup {
break
}
s.pipeline.SetTask(t)
now := time.Now()
nodeLess := func(a *NodeInfo, b *NodeInfo) bool {
// If either node has at least maxFailures recent failures,
// that's the deciding factor.
recentFailuresA := a.countRecentFailures(now, t.ServiceID)
recentFailuresB := b.countRecentFailures(now, t.ServiceID)
if recentFailuresA >= maxFailures || recentFailuresB >= maxFailures {
if recentFailuresA > recentFailuresB {
return false
}
if recentFailuresB > recentFailuresA {
return true
}
}
tasksByServiceA := a.DesiredRunningTasksCountByService[t.ServiceID]
tasksByServiceB := b.DesiredRunningTasksCountByService[t.ServiceID]
if tasksByServiceA < tasksByServiceB {
return true
}
if tasksByServiceA > tasksByServiceB {
return false
}
// Total number of tasks breaks ties.
return a.DesiredRunningTasksCount < b.DesiredRunningTasksCount
}
var prefs []*api.PlacementPreference
if t.Spec.Placement != nil {
prefs = t.Spec.Placement.Preferences
}
tree := s.nodeSet.tree(t.ServiceID, prefs, len(taskGroup), s.pipeline.Process, nodeLess)
s.scheduleNTasksOnSubtree(ctx, len(taskGroup), taskGroup, &tree, schedulingDecisions, nodeLess)
if len(taskGroup) != 0 {
s.noSuitableNode(ctx, taskGroup, schedulingDecisions)
}
}
func (s *Scheduler) scheduleNTasksOnSubtree(ctx context.Context, n int, taskGroup map[string]*api.Task, tree *decisionTree, schedulingDecisions map[string]schedulingDecision, nodeLess func(a *NodeInfo, b *NodeInfo) bool) int {
if tree.next == nil {
nodes := tree.orderedNodes(s.pipeline.Process, nodeLess)
if len(nodes) == 0 {
return 0
}
return s.scheduleNTasksOnNodes(ctx, n, taskGroup, nodes, schedulingDecisions, nodeLess)
}
// Walk the tree and figure out how the tasks should be split at each
// level.
tasksScheduled := 0
tasksInUsableBranches := tree.tasks
var noRoom map[*decisionTree]struct{}
// Try to make branches even until either all branches are
// full, or all tasks have been scheduled.
for tasksScheduled != n && len(noRoom) != len(tree.next) {
desiredTasksPerBranch := (tasksInUsableBranches + n - tasksScheduled) / (len(tree.next) - len(noRoom))
remainder := (tasksInUsableBranches + n - tasksScheduled) % (len(tree.next) - len(noRoom))
for _, subtree := range tree.next {
if noRoom != nil {
if _, ok := noRoom[subtree]; ok {
continue
}
}
subtreeTasks := subtree.tasks
if subtreeTasks < desiredTasksPerBranch || (subtreeTasks == desiredTasksPerBranch && remainder > 0) {
tasksToAssign := desiredTasksPerBranch - subtreeTasks
if remainder > 0 {
tasksToAssign++
}
res := s.scheduleNTasksOnSubtree(ctx, tasksToAssign, taskGroup, subtree, schedulingDecisions, nodeLess)
if res < tasksToAssign {
if noRoom == nil {
noRoom = make(map[*decisionTree]struct{})
}
noRoom[subtree] = struct{}{}
tasksInUsableBranches -= subtreeTasks
} else if remainder > 0 {
remainder--
}
tasksScheduled += res
}
}
}
return tasksScheduled
}
func (s *Scheduler) scheduleNTasksOnNodes(ctx context.Context, n int, taskGroup map[string]*api.Task, nodes []NodeInfo, schedulingDecisions map[string]schedulingDecision, nodeLess func(a *NodeInfo, b *NodeInfo) bool) int {
tasksScheduled := 0
failedConstraints := make(map[int]bool) // key is index in nodes slice
nodeIter := 0
nodeCount := len(nodes)
for taskID, t := range taskGroup {
node := &nodes[nodeIter%nodeCount]
log.G(ctx).WithField("task.id", t.ID).Debugf("assigning to node %s", node.ID)
newT := *t
newT.NodeID = node.ID
newT.Status = api.TaskStatus{
State: api.TaskStateAssigned,
Timestamp: ptypes.MustTimestampProto(time.Now()),
Message: "scheduler assigned task to node",
}
s.allTasks[t.ID] = &newT
nodeInfo, err := s.nodeSet.nodeInfo(node.ID)
if err == nil && nodeInfo.addTask(&newT) {
s.nodeSet.updateNode(nodeInfo)
nodes[nodeIter%nodeCount] = nodeInfo
}
schedulingDecisions[taskID] = schedulingDecision{old: t, new: &newT}
delete(taskGroup, taskID)
tasksScheduled++
if tasksScheduled == n {
return tasksScheduled
}
if nodeIter+1 < nodeCount {
// First pass fills the nodes until they have the same
// number of tasks from this service.
nextNode := nodes[(nodeIter+1)%nodeCount]
if nodeLess(&nextNode, &nodeInfo) {
nodeIter++
}
} else {
// In later passes, we just assign one task at a time
// to each node that still meets the constraints.
nodeIter++
}
origNodeIter := nodeIter
for failedConstraints[nodeIter%nodeCount] || !s.pipeline.Process(&nodes[nodeIter%nodeCount]) {
failedConstraints[nodeIter%nodeCount] = true
nodeIter++
if nodeIter-origNodeIter == nodeCount {
// None of the nodes meet the constraints anymore.
return tasksScheduled
}
}
}
return tasksScheduled
}
func (s *Scheduler) noSuitableNode(ctx context.Context, taskGroup map[string]*api.Task, schedulingDecisions map[string]schedulingDecision) {
explanation := s.pipeline.Explain()
for _, t := range taskGroup {
log.G(ctx).WithField("task.id", t.ID).Debug("no suitable node available for task")
newT := *t
newT.Status.Timestamp = ptypes.MustTimestampProto(time.Now())
if explanation != "" {
newT.Status.Message = "no suitable node (" + explanation + ")"
} else {
newT.Status.Message = "no suitable node"
}
s.allTasks[t.ID] = &newT
schedulingDecisions[t.ID] = schedulingDecision{old: t, new: &newT}
s.enqueue(&newT)
}
}
func (s *Scheduler) buildNodeSet(tx store.ReadTx, tasksByNode map[string]map[string]*api.Task) error {
nodes, err := store.FindNodes(tx, store.All)
if err != nil {
return err
}
s.nodeSet.alloc(len(nodes))
for _, n := range nodes {
var resources api.Resources
if n.Description != nil && n.Description.Resources != nil {
resources = *n.Description.Resources
}
s.nodeSet.addOrUpdateNode(newNodeInfo(n, tasksByNode[n.ID], resources))
}
return nil
}