-
Notifications
You must be signed in to change notification settings - Fork 134
/
simulator.go
867 lines (818 loc) · 29.4 KB
/
simulator.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
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
package simulator
import (
"container/heap"
"fmt"
"math/rand"
"sync/atomic"
"time"
"github.com/gogo/protobuf/proto"
"github.com/oklog/ulid"
"github.com/pkg/errors"
"golang.org/x/exp/maps"
"golang.org/x/exp/slices"
"golang.org/x/time/rate"
v1 "k8s.io/api/core/v1"
"github.com/armadaproject/armada/internal/armada/configuration"
"github.com/armadaproject/armada/internal/common/armadacontext"
"github.com/armadaproject/armada/internal/common/logging"
armadaslices "github.com/armadaproject/armada/internal/common/slices"
"github.com/armadaproject/armada/internal/common/types"
"github.com/armadaproject/armada/internal/common/util"
"github.com/armadaproject/armada/internal/scheduler"
schedulerconstraints "github.com/armadaproject/armada/internal/scheduler/constraints"
schedulercontext "github.com/armadaproject/armada/internal/scheduler/context"
"github.com/armadaproject/armada/internal/scheduler/fairness"
"github.com/armadaproject/armada/internal/scheduler/jobdb"
"github.com/armadaproject/armada/internal/scheduler/nodedb"
schedulerobjects "github.com/armadaproject/armada/internal/scheduler/schedulerobjects"
"github.com/armadaproject/armada/internal/scheduleringester"
"github.com/armadaproject/armada/pkg/armadaevents"
)
// Simulator captures the parameters and state of the Armada simulator.
type Simulator struct {
ClusterSpec *ClusterSpec
WorkloadSpec *WorkloadSpec
schedulingConfig configuration.SchedulingConfig
// Map from jobId to the jobTemplate from which the job was created.
jobTemplateByJobId map[string]*JobTemplate
// Map from job template ids to slices of templates depending on those ids.
jobTemplatesByDependencyIds map[string]map[string]*JobTemplate
// Map from job template id to jobTemplate for templates for which all jobs have not yet succeeded.
activeJobTemplatesById map[string]*JobTemplate
// The JobDb stores all jobs that have yet to terminate.
jobDb *jobdb.JobDb
// Map from node id to the pool to which the node belongs.
poolByNodeId map[string]string
// Separate nodeDb per pool and executorGroup.
nodeDbByPoolAndExecutorGroup map[string][]*nodedb.NodeDb
// Map from executor name to the nodeDb to which it belongs.
nodeDbByExecutorName map[string]*nodedb.NodeDb
// Allocation by pool for each queue and priority class.
// Stored across invocations of the scheduler.
allocationByPoolAndQueueAndPriorityClass map[string]map[string]schedulerobjects.QuantityByTAndResourceType[string]
// Total resources across all executorGroups for each pool.
totalResourcesByPool map[string]schedulerobjects.ResourceList
// Indicates whether a job has been submitted or terminated since the last scheduling round.
shouldSchedule bool
// Current simulated time.
time time.Time
// Sequence number of the next event to be published.
sequenceNumber int
// Events stored in a priority queue ordered first by timestamp and second by sequence number.
eventLog EventLog
// Simulated events are emitted on these event channels.
// Create a channel by calling s.StateTransitions() before running the simulator.
stateTransitionChannels []chan StateTransition
// Global job scheduling rate-limiter.
limiter *rate.Limiter
// Per-queue job scheduling rate-limiters.
limiterByQueue map[string]*rate.Limiter
// Used to generate random numbers from a chosen seed.
rand *rand.Rand
// Used to ensure each job is given a unique time stamp.
logicalJobCreatedTimestamp atomic.Int64
// If true, scheduler logs are omitted.
// This since the logs are very verbose when scheduling large numbers of jobs.
SuppressSchedulerLogs bool
}
type StateTransition struct {
Jobs []*jobdb.Job
EventSequence *armadaevents.EventSequence
}
func NewSimulator(clusterSpec *ClusterSpec, workloadSpec *WorkloadSpec, schedulingConfig configuration.SchedulingConfig) (*Simulator, error) {
// TODO: Move clone to caller?
// Copy specs to avoid concurrent mutation.
clusterSpec = proto.Clone(clusterSpec).(*ClusterSpec)
workloadSpec = proto.Clone(workloadSpec).(*WorkloadSpec)
initialiseClusterSpec(clusterSpec)
initialiseWorkloadSpec(workloadSpec)
if err := validateClusterSpec(clusterSpec); err != nil {
return nil, err
}
if err := validateWorkloadSpec(workloadSpec); err != nil {
return nil, err
}
jobDb := jobdb.NewJobDb(
schedulingConfig.Preemption.PriorityClasses,
schedulingConfig.Preemption.DefaultPriorityClass,
1024,
)
randomSeed := workloadSpec.RandomSeed
if randomSeed == 0 {
// Seed the RNG using the local time if no explic random seed is provided.
randomSeed = time.Now().Unix()
}
s := &Simulator{
ClusterSpec: clusterSpec,
WorkloadSpec: workloadSpec,
schedulingConfig: schedulingConfig,
jobTemplateByJobId: make(map[string]*JobTemplate),
jobTemplatesByDependencyIds: make(map[string]map[string]*JobTemplate),
activeJobTemplatesById: make(map[string]*JobTemplate),
jobDb: jobDb,
nodeDbByPoolAndExecutorGroup: make(map[string][]*nodedb.NodeDb),
poolByNodeId: make(map[string]string),
nodeDbByExecutorName: make(map[string]*nodedb.NodeDb),
allocationByPoolAndQueueAndPriorityClass: make(map[string]map[string]schedulerobjects.QuantityByTAndResourceType[string]),
totalResourcesByPool: make(map[string]schedulerobjects.ResourceList),
limiter: rate.NewLimiter(
rate.Limit(schedulingConfig.MaximumSchedulingRate),
schedulingConfig.MaximumSchedulingBurst,
),
limiterByQueue: make(map[string]*rate.Limiter),
rand: rand.New(rand.NewSource(randomSeed)),
}
jobDb.SetClock(s)
s.limiter.SetBurstAt(s.time, schedulingConfig.MaximumSchedulingBurst)
if err := s.setupClusters(); err != nil {
return nil, err
}
if err := s.bootstrapWorkload(); err != nil {
return nil, err
}
return s, nil
}
func (s *Simulator) Now() time.Time {
return s.time
}
func (s *Simulator) Since(t time.Time) time.Duration {
return s.Now().Sub(t)
}
// Run runs the scheduler until all jobs have finished successfully.
func (s *Simulator) Run(ctx *armadacontext.Context) error {
defer func() {
for _, c := range s.stateTransitionChannels {
close(c)
}
}()
// Bootstrap the simulator by pushing an event that triggers a scheduler run.
s.pushScheduleEvent(s.time)
// Then run the scheduler until all jobs have completed.
for s.eventLog.Len() > 0 {
select {
case <-ctx.Done():
return ctx.Err()
default:
event := heap.Pop(&s.eventLog).(Event)
if err := s.handleSimulatorEvent(ctx, event); err != nil {
return err
}
}
}
return nil
}
// StateTransitions returns a channel on which all simulated events are sent.
// This function must be called before *Simulator.Run.
func (s *Simulator) StateTransitions() <-chan StateTransition {
c := make(chan StateTransition, 128)
s.stateTransitionChannels = append(s.stateTransitionChannels, c)
return c
}
func validateClusterSpec(clusterSpec *ClusterSpec) error {
poolNames := util.Map(clusterSpec.Pools, func(pool *Pool) string { return pool.Name })
if !slices.Equal(poolNames, armadaslices.Unique(poolNames)) {
return errors.Errorf("duplicate pool name: %v", poolNames)
}
executorNames := make([]string, 0)
for _, pool := range clusterSpec.Pools {
for _, executorGroup := range pool.ClusterGroups {
for _, executor := range executorGroup.Clusters {
executorNames = append(executorNames, executor.Name)
}
}
}
if !slices.Equal(executorNames, armadaslices.Unique(executorNames)) {
return errors.Errorf("duplicate executor name: %v", executorNames)
}
return nil
}
func validateWorkloadSpec(workloadSpec *WorkloadSpec) error {
queueNames := util.Map(workloadSpec.Queues, func(queue *Queue) string { return queue.Name })
if !slices.Equal(queueNames, armadaslices.Unique(queueNames)) {
return errors.Errorf("duplicate queue name: %v", queueNames)
}
jobTemplateIdSlices := util.Map(workloadSpec.Queues, func(queue *Queue) []string {
return util.Map(queue.JobTemplates, func(template *JobTemplate) string { return template.Id })
})
jobTemplateIds := make([]string, 0)
for _, singleQueueTemplateIds := range jobTemplateIdSlices {
jobTemplateIds = append(jobTemplateIds, singleQueueTemplateIds...)
}
if !slices.Equal(jobTemplateIds, armadaslices.Unique(jobTemplateIds)) {
return errors.Errorf("duplicate job template ids: %v", jobTemplateIds)
}
return nil
}
func (s *Simulator) setupClusters() error {
for _, pool := range s.ClusterSpec.Pools {
totalResourcesForPool := schedulerobjects.ResourceList{}
for executorGroupIndex, executorGroup := range pool.ClusterGroups {
nodeDb, err := nodedb.NewNodeDb(
s.schedulingConfig.Preemption.PriorityClasses,
s.schedulingConfig.MaxExtraNodesToConsider,
s.schedulingConfig.IndexedResources,
s.schedulingConfig.IndexedTaints,
s.schedulingConfig.IndexedNodeLabels,
s.schedulingConfig.WellKnownNodeTypes,
)
if err != nil {
return err
}
for executorIndex, executor := range executorGroup.Clusters {
executorName := fmt.Sprintf("%s-%d-%d", pool.Name, executorGroupIndex, executorIndex)
s.nodeDbByExecutorName[executorName] = nodeDb
for nodeTemplateIndex, nodeTemplate := range executor.NodeTemplates {
for i := 0; i < int(nodeTemplate.Number); i++ {
nodeId := fmt.Sprintf("%s-%d-%d-%d-%d", pool.Name, executorGroupIndex, executorIndex, nodeTemplateIndex, i)
node := &schedulerobjects.Node{
Id: nodeId,
Name: nodeId,
Executor: executorName,
Taints: slices.Clone(nodeTemplate.Taints),
Labels: maps.Clone(nodeTemplate.Labels),
TotalResources: nodeTemplate.TotalResources.DeepCopy(),
AllocatableByPriorityAndResource: schedulerobjects.NewAllocatableByPriorityAndResourceType(
types.AllowedPriorities(s.schedulingConfig.Preemption.PriorityClasses),
nodeTemplate.TotalResources,
),
}
txn := nodeDb.Txn(true)
if err := nodeDb.CreateAndInsertWithApiJobsWithTxn(txn, nil, node); err != nil {
txn.Abort()
return err
}
txn.Commit()
s.poolByNodeId[nodeId] = pool.Name
}
}
}
s.nodeDbByPoolAndExecutorGroup[pool.Name] = append(s.nodeDbByPoolAndExecutorGroup[pool.Name], nodeDb)
totalResourcesForPool.Add(nodeDb.TotalResources())
}
s.totalResourcesByPool[pool.Name] = totalResourcesForPool
}
return nil
}
func (s *Simulator) bootstrapWorkload() error {
// Mark all jobTemplates as active.
for _, queue := range s.WorkloadSpec.Queues {
for _, jobTemplate := range queue.JobTemplates {
s.activeJobTemplatesById[jobTemplate.Id] = jobTemplate
}
}
// Publish submitJob messages for all jobTemplates without dependencies.
for _, queue := range s.WorkloadSpec.Queues {
for _, jobTemplate := range queue.JobTemplates {
if len(jobTemplate.Dependencies) > 0 {
continue
}
eventSequence := &armadaevents.EventSequence{
Queue: queue.Name,
JobSetName: jobTemplate.JobSet,
}
for k := 0; k < int(jobTemplate.Number); k++ {
if len(jobTemplate.Dependencies) > 0 {
continue
}
jobId := util.ULID()
eventSequence.Events = append(
eventSequence.Events,
&armadaevents.EventSequence_Event{
Created: pointer(s.time.Add(jobTemplate.EarliestSubmitTime)),
Event: &armadaevents.EventSequence_Event_SubmitJob{
SubmitJob: submitJobFromJobTemplate(jobId, jobTemplate),
},
},
)
s.jobTemplateByJobId[jobId.String()] = jobTemplate
}
if len(eventSequence.Events) > 0 {
s.pushEventSequence(eventSequence)
}
}
}
// Setup the jobTemplate dependency map.
for _, queue := range s.WorkloadSpec.Queues {
for _, jobTemplate := range queue.JobTemplates {
for _, dependencyJobTemplateId := range jobTemplate.Dependencies {
dependencyJobTemplate, ok := s.activeJobTemplatesById[dependencyJobTemplateId]
if !ok {
return errors.Errorf(
"jobTemplate %s depends on jobTemplate %s, which does not exist",
jobTemplate.Id, dependencyJobTemplate.Id,
)
}
m := s.jobTemplatesByDependencyIds[dependencyJobTemplateId]
if m == nil {
m = make(map[string]*JobTemplate)
s.jobTemplatesByDependencyIds[dependencyJobTemplateId] = m
}
m[jobTemplate.Id] = jobTemplate
}
}
}
return nil
}
func submitJobFromJobTemplate(jobId ulid.ULID, jobTemplate *JobTemplate) *armadaevents.SubmitJob {
return &armadaevents.SubmitJob{
JobId: armadaevents.ProtoUuidFromUlid(jobId),
Priority: jobTemplate.QueuePriority,
MainObject: &armadaevents.KubernetesMainObject{
ObjectMeta: &armadaevents.ObjectMeta{
Annotations: jobTemplate.Requirements.Annotations,
},
Object: &armadaevents.KubernetesMainObject_PodSpec{
PodSpec: &armadaevents.PodSpecWithAvoidList{
PodSpec: &v1.PodSpec{
NodeSelector: jobTemplate.Requirements.NodeSelector,
Affinity: jobTemplate.Requirements.Affinity,
Tolerations: jobTemplate.Requirements.Tolerations,
PriorityClassName: jobTemplate.PriorityClassName,
Containers: []v1.Container{
{
Resources: jobTemplate.Requirements.ResourceRequirements,
},
},
},
},
},
},
}
}
func (s *Simulator) pushEventSequence(eventSequence *armadaevents.EventSequence) {
if len(eventSequence.Events) == 0 {
return
}
heap.Push(
&s.eventLog,
Event{
// We assume that all events in the sequence have the same Created time.
time: *eventSequence.Events[0].Created,
sequenceNumber: s.sequenceNumber,
eventSequenceOrScheduleEvent: eventSequence,
},
)
s.sequenceNumber++
}
func (s *Simulator) pushScheduleEvent(time time.Time) {
heap.Push(
&s.eventLog,
Event{
time: time,
sequenceNumber: s.sequenceNumber,
eventSequenceOrScheduleEvent: scheduleEvent{},
},
)
s.sequenceNumber++
}
func (s *Simulator) handleSimulatorEvent(ctx *armadacontext.Context, event Event) error {
s.time = event.time
ctx = armadacontext.New(ctx.Context, ctx.FieldLogger.WithField("simulated time", event.time))
switch e := event.eventSequenceOrScheduleEvent.(type) {
case *armadaevents.EventSequence:
if err := s.handleEventSequence(ctx, e); err != nil {
return err
}
case scheduleEvent:
if err := s.handleScheduleEvent(ctx); err != nil {
return err
}
}
return nil
}
func (s *Simulator) handleScheduleEvent(ctx *armadacontext.Context) error {
// Schedule the next run of the scheduler, unless there are no more active jobTemplates.
// TODO: Make timeout configurable.
if len(s.activeJobTemplatesById) > 0 {
s.pushScheduleEvent(s.time.Add(10 * time.Second))
}
if !s.shouldSchedule {
return nil
}
var eventSequences []*armadaevents.EventSequence
txn := s.jobDb.WriteTxn()
defer txn.Abort()
for _, pool := range s.ClusterSpec.Pools {
for i := range pool.ClusterGroups {
nodeDb := s.nodeDbByPoolAndExecutorGroup[pool.Name][i]
if err := nodeDb.Reset(); err != nil {
return err
}
totalResources := s.totalResourcesByPool[pool.Name]
fairnessCostProvider, err := fairness.NewDominantResourceFairness(
totalResources,
s.schedulingConfig.DominantResourceFairnessResourcesToConsider,
)
if err != nil {
return err
}
sctx := schedulercontext.NewSchedulingContext(
fmt.Sprintf("%s-%d", pool.Name, i),
pool.Name,
s.schedulingConfig.Preemption.PriorityClasses,
s.schedulingConfig.Preemption.DefaultPriorityClass,
fairnessCostProvider,
s.limiter,
totalResources,
)
sctx.Started = s.time
for _, queue := range s.WorkloadSpec.Queues {
limiter, ok := s.limiterByQueue[queue.Name]
if !ok {
limiter = rate.NewLimiter(
rate.Limit(s.schedulingConfig.MaximumPerQueueSchedulingRate),
s.schedulingConfig.MaximumPerQueueSchedulingBurst,
)
limiter.SetBurstAt(s.time, s.schedulingConfig.MaximumPerQueueSchedulingBurst)
s.limiterByQueue[queue.Name] = limiter
}
err := sctx.AddQueueSchedulingContext(
queue.Name,
queue.Weight,
s.allocationByPoolAndQueueAndPriorityClass[pool.Name][queue.Name],
limiter,
)
if err != nil {
return err
}
}
constraints := schedulerconstraints.NewSchedulingConstraints(
pool.Name,
totalResources,
// Minimum job size not used for simulation; use taints/tolerations instead.
schedulerobjects.ResourceList{},
s.schedulingConfig,
nil,
)
sch := scheduler.NewPreemptingQueueScheduler(
sctx,
constraints,
s.schedulingConfig.Preemption.NodeEvictionProbability,
s.schedulingConfig.Preemption.NodeOversubscriptionEvictionProbability,
s.schedulingConfig.Preemption.ProtectedFractionOfFairShare,
scheduler.NewSchedulerJobRepositoryAdapter(txn),
nodeDb,
// TODO: Necessary to support partial eviction.
nil,
nil,
nil,
)
schedulerCtx := ctx
if s.SuppressSchedulerLogs {
schedulerCtx = &armadacontext.Context{
Context: ctx.Context,
FieldLogger: logging.NullLogger,
}
}
result, err := sch.Schedule(schedulerCtx)
if err != nil {
return err
}
// Update jobDb to reflect the decisions by the scheduler.
// Sort jobs to ensure deterministic event ordering.
preemptedJobs := scheduler.PreemptedJobsFromSchedulerResult[*jobdb.Job](result)
scheduledJobs := slices.Clone(result.ScheduledJobs)
failedJobs := scheduler.FailedJobsFromSchedulerResult[*jobdb.Job](result)
lessJob := func(a, b *jobdb.Job) int {
if a.Queue() < b.Queue() {
return -1
} else if a.Queue() > b.Queue() {
return 1
}
if a.Id() < b.Id() {
return -1
} else if a.Id() > b.Id() {
return 1
}
return 0
}
slices.SortFunc(preemptedJobs, lessJob)
slices.SortFunc(scheduledJobs, func(a, b *schedulercontext.JobSchedulingContext) int {
return lessJob(a.Job.(*jobdb.Job), b.Job.(*jobdb.Job))
})
slices.SortFunc(failedJobs, lessJob)
for i, job := range preemptedJobs {
if run := job.LatestRun(); run != nil {
job = job.WithUpdatedRun(run.WithFailed(true))
} else {
return errors.Errorf("attempting to preempt job %s with no associated runs", job.Id())
}
preemptedJobs[i] = job.WithQueued(false).WithFailed(true)
}
for i, jctx := range scheduledJobs {
job := jctx.Job.(*jobdb.Job)
nodeId := result.NodeIdByJobId[job.GetId()]
if nodeId == "" {
return errors.Errorf("job %s not mapped to a node", job.GetId())
}
if node, err := nodeDb.GetNode(nodeId); err != nil {
return err
} else {
priority, ok := nodeDb.GetScheduledAtPriority(job.Id())
if !ok {
return errors.Errorf("job %s not mapped to a priority", job.Id())
}
scheduledJobs[i].Job = job.WithQueued(false).WithNewRun(node.Executor, node.Id, node.Name, priority)
}
}
for i, job := range failedJobs {
if run := job.LatestRun(); run != nil {
job = job.WithUpdatedRun(run.WithFailed(true))
}
failedJobs[i] = job.WithQueued(false).WithFailed(true)
}
if err := txn.Upsert(preemptedJobs); err != nil {
return err
}
if err := txn.Upsert(util.Map(scheduledJobs, func(jctx *schedulercontext.JobSchedulingContext) *jobdb.Job { return jctx.Job.(*jobdb.Job) })); err != nil {
return err
}
if err := txn.Upsert(failedJobs); err != nil {
return err
}
// Update allocation.
s.allocationByPoolAndQueueAndPriorityClass[pool.Name] = sctx.AllocatedByQueueAndPriority()
// Generate eventSequences.
// TODO: Add time taken to run the scheduler to s.time.
eventSequences, err = scheduler.AppendEventSequencesFromPreemptedJobs(eventSequences, preemptedJobs, s.time)
if err != nil {
return err
}
eventSequences, err = scheduler.AppendEventSequencesFromScheduledJobs(eventSequences, scheduledJobs, make(map[string]map[string]string))
if err != nil {
return err
}
eventSequences, err = scheduler.AppendEventSequencesFromUnschedulableJobs(eventSequences, failedJobs, s.time)
if err != nil {
return err
}
// Update event timestamps to be consistent with simulated time.
t := s.time
for _, eventSequence := range eventSequences {
for _, event := range eventSequence.Events {
event.Created = &t
}
}
// If nothing changed, we're in steady state and can safely skip scheduling until something external has changed.
// Do this only if a non-zero amount of time has passed.
if !s.time.Equal(time.Time{}) && len(result.ScheduledJobs) == 0 && len(result.PreemptedJobs) == 0 && len(result.FailedJobs) == 0 {
s.shouldSchedule = false
}
}
}
txn.Commit()
// Publish simulator events.
for _, eventSequence := range eventSequences {
s.pushEventSequence(eventSequence)
}
return nil
}
// TODO: Write events to disk unless they should be discarded.
func (s *Simulator) handleEventSequence(ctx *armadacontext.Context, es *armadaevents.EventSequence) error {
txn := s.jobDb.WriteTxn()
defer txn.Abort()
eventsToPublish := make([]*armadaevents.EventSequence_Event, 0, len(es.Events))
jobs := make([]*jobdb.Job, len(es.Events))
for i, event := range es.Events {
var err error
var shouldPublish bool
switch eventType := event.GetEvent().(type) {
case *armadaevents.EventSequence_Event_SubmitJob:
s.shouldSchedule = true
jobs[i], shouldPublish, err = s.handleSubmitJob(txn, event.GetSubmitJob(), *event.Created, es)
case *armadaevents.EventSequence_Event_JobRunLeased:
jobs[i], shouldPublish, err = s.handleJobRunLeased(txn, event.GetJobRunLeased())
case *armadaevents.EventSequence_Event_JobSucceeded:
s.shouldSchedule = true
jobs[i], shouldPublish, err = s.handleJobSucceeded(txn, event.GetJobSucceeded())
case *armadaevents.EventSequence_Event_JobRunPreempted:
s.shouldSchedule = true
jobs[i], shouldPublish, err = s.handleJobRunPreempted(txn, event.GetJobRunPreempted())
case *armadaevents.EventSequence_Event_JobRunErrors:
for _, e := range event.GetJobRunErrors().Errors {
if e.GetJobRunPreemptedError() == nil {
return errors.Errorf("received unexpected JobRunErrors reason: %T", e.Reason)
}
}
case *armadaevents.EventSequence_Event_JobErrors:
for _, e := range event.GetJobErrors().Errors {
if e.GetJobRunPreemptedError() == nil {
return errors.Errorf("received unexpected JobErrors reason: %T", e.Reason)
}
}
default:
// This is an event type we haven't considered
return errors.Errorf("received unknown event type %T", eventType)
}
if err != nil {
return err
}
if shouldPublish {
eventsToPublish = append(eventsToPublish, event)
}
}
txn.Commit()
es.Events = eventsToPublish
if len(es.Events) > 0 {
stateTransition := StateTransition{
Jobs: jobs,
EventSequence: es,
}
for _, c := range s.stateTransitionChannels {
c <- stateTransition
}
}
return nil
}
func (s *Simulator) handleSubmitJob(txn *jobdb.Txn, e *armadaevents.SubmitJob, time time.Time, eventSequence *armadaevents.EventSequence) (*jobdb.Job, bool, error) {
schedulingInfo, err := scheduleringester.SchedulingInfoFromSubmitJob(e, time, s.schedulingConfig.Preemption.PriorityClasses)
if err != nil {
return nil, false, err
}
job := s.jobDb.NewJob(
armadaevents.UlidFromProtoUuid(e.JobId).String(),
eventSequence.JobSetName,
eventSequence.Queue,
e.Priority,
schedulingInfo,
true,
0,
false,
false,
false,
s.logicalJobCreatedTimestamp.Add(1),
)
if err := txn.Upsert([]*jobdb.Job{job}); err != nil {
return nil, false, err
}
return job, true, nil
}
func (s *Simulator) handleJobRunLeased(txn *jobdb.Txn, e *armadaevents.JobRunLeased) (*jobdb.Job, bool, error) {
jobId := armadaevents.UlidFromProtoUuid(e.JobId).String()
job := txn.GetById(jobId)
jobTemplate := s.jobTemplateByJobId[jobId]
if jobTemplate == nil {
return nil, false, errors.Errorf("no jobTemplate associated with job %s", jobId)
}
jobSuccessTime := s.time
jobSuccessTime = jobSuccessTime.Add(s.generateRandomShiftedExponentialDuration(s.ClusterSpec.PendingDelayDistribution))
jobSuccessTime = jobSuccessTime.Add(s.generateRandomShiftedExponentialDuration(jobTemplate.RuntimeDistribution))
s.pushEventSequence(
&armadaevents.EventSequence{
Queue: job.Queue(),
JobSetName: job.Jobset(),
Events: []*armadaevents.EventSequence_Event{
{
Created: &jobSuccessTime,
Event: &armadaevents.EventSequence_Event_JobSucceeded{
JobSucceeded: &armadaevents.JobSucceeded{
JobId: e.JobId,
},
},
},
},
},
)
updatedJob := job.WithUpdatedRun(job.LatestRun().WithRunning(true))
if err := txn.Upsert([]*jobdb.Job{updatedJob}); err != nil {
return nil, false, err
}
return updatedJob, true, nil
}
func (s *Simulator) generateRandomShiftedExponentialDuration(rv ShiftedExponential) time.Duration {
return generateRandomShiftedExponentialDuration(s.rand, rv)
}
func generateRandomShiftedExponentialDuration(r *rand.Rand, rv ShiftedExponential) time.Duration {
if rv.TailMean == 0 {
return rv.Minimum
} else {
return rv.Minimum + time.Duration(r.ExpFloat64()*float64(rv.TailMean))
}
}
func (s *Simulator) handleJobSucceeded(txn *jobdb.Txn, e *armadaevents.JobSucceeded) (*jobdb.Job, bool, error) {
jobId := armadaevents.UlidFromProtoUuid(e.JobId).String()
job := txn.GetById(jobId)
if job == nil || job.InTerminalState() {
// Job already terminated; nothing more to do.
return nil, false, nil
}
if err := txn.BatchDelete([]string{jobId}); err != nil {
return nil, false, err
}
// Subtract the allocation of this job from the queue allocation.
run := job.LatestRun()
pool := s.poolByNodeId[run.NodeId()]
s.allocationByPoolAndQueueAndPriorityClass[pool][job.Queue()].SubV1ResourceList(
job.GetPriorityClassName(),
job.GetResourceRequirements().Requests,
)
// Unbind the job from the node on which it was scheduled.
if err := s.unbindRunningJob(job); err != nil {
return nil, false, errors.WithMessagef(err, "failed to unbind job %s", job.Id())
}
// Increase the successful job count for this jobTemplate.
// If all jobs created from this template have succeeded, update dependent templates
// and submit any templates for which this was the last dependency.
jobTemplate := s.jobTemplateByJobId[job.GetId()]
jobTemplate.NumberSuccessful++
if jobTemplate.Number == jobTemplate.NumberSuccessful {
delete(s.activeJobTemplatesById, jobTemplate.Id)
for _, dependentJobTemplate := range s.jobTemplatesByDependencyIds[jobTemplate.Id] {
i := slices.Index(dependentJobTemplate.Dependencies, jobTemplate.Id)
dependentJobTemplate.Dependencies = slices.Delete(dependentJobTemplate.Dependencies, i, i+1)
if len(dependentJobTemplate.Dependencies) > 0 {
continue
}
eventSequence := &armadaevents.EventSequence{
Queue: dependentJobTemplate.Queue,
JobSetName: dependentJobTemplate.JobSet,
}
for k := 0; k < int(dependentJobTemplate.Number); k++ {
jobId := util.ULID()
eventSequence.Events = append(
eventSequence.Events,
&armadaevents.EventSequence_Event{
// EarliestSubmitTimeFromDependencyCompletion must be positive
Created: pointer(maxTime(time.Time{}.Add(dependentJobTemplate.EarliestSubmitTime), s.time.Add(dependentJobTemplate.EarliestSubmitTimeFromDependencyCompletion))),
Event: &armadaevents.EventSequence_Event_SubmitJob{
SubmitJob: submitJobFromJobTemplate(jobId, dependentJobTemplate),
},
},
)
s.jobTemplateByJobId[jobId.String()] = dependentJobTemplate
}
if len(eventSequence.Events) > 0 {
s.pushEventSequence(eventSequence)
}
}
delete(s.jobTemplatesByDependencyIds, jobTemplate.Id)
}
return job.WithSucceeded(true).WithUpdatedRun(run.WithRunning(false).WithSucceeded(true)), true, nil
}
func (s *Simulator) unbindRunningJob(job *jobdb.Job) error {
if job.InTerminalState() {
return errors.Errorf("job %s has terminated", job.Id())
}
run := job.LatestRun()
if run == nil {
return errors.Errorf("job %s has no runs associated with it", job.Id())
}
if run.Executor() == "" {
return errors.Errorf("empty executor for run %s of job %s", run.Id(), job.Id())
}
if run.NodeId() == "" {
return errors.Errorf("empty nodeId for run %s of job %s", run.Id(), job.Id())
}
nodeDb := s.nodeDbByExecutorName[run.Executor()]
node, err := nodeDb.GetNode(run.NodeId())
if err != nil {
return err
} else if node == nil {
return errors.Errorf("node %s not found", run.NodeId())
}
node, err = nodeDb.UnbindJobFromNode(s.schedulingConfig.Preemption.PriorityClasses, job, node)
if err != nil {
return err
}
if err := nodeDb.Upsert(node); err != nil {
return err
}
return nil
}
func (s *Simulator) handleJobRunPreempted(txn *jobdb.Txn, e *armadaevents.JobRunPreempted) (*jobdb.Job, bool, error) {
jobId := armadaevents.UlidFromProtoUuid(e.PreemptedJobId).String()
job := txn.GetById(jobId)
// Submit a retry for this job.
jobTemplate := s.jobTemplateByJobId[job.GetId()]
retryJobId := util.ULID()
resubmitTime := s.time.Add(s.generateRandomShiftedExponentialDuration(s.ClusterSpec.WorkflowManagerDelayDistribution))
s.pushEventSequence(
&armadaevents.EventSequence{
Queue: job.Queue(),
JobSetName: job.Jobset(),
Events: []*armadaevents.EventSequence_Event{
{
Created: &resubmitTime,
Event: &armadaevents.EventSequence_Event_SubmitJob{
SubmitJob: submitJobFromJobTemplate(retryJobId, jobTemplate),
},
},
},
},
)
s.jobTemplateByJobId[retryJobId.String()] = jobTemplate
updatedJob := job.WithUpdatedRun(job.LatestRun().WithReturned(true))
if err := txn.Upsert([]*jobdb.Job{updatedJob}); err != nil {
return nil, false, err
}
return updatedJob, true, nil
}
func maxTime(a, b time.Time) time.Time {
if a.Before(b) {
return b
}
return a
}
func pointer[T any](t T) *T {
return &t
}