lib/dispatchcloud/scheduler/scheduler.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 // Package scheduler uses a resizable worker pool to execute
   6 // containers in priority order.
   7 package scheduler
   8
   9 import (
  10         "context"
  11         "sync"
  12         "time"
  13
  14         "git.arvados.org/arvados.git/sdk/go/arvados"
  15         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  16         "github.com/prometheus/client_golang/prometheus"
  17         "github.com/sirupsen/logrus"
  18 )
  19
  20 // A Scheduler maps queued containers onto unallocated workers in
  21 // priority order, creating new workers if needed. It locks containers
  22 // that can be mapped onto existing/pending workers, and starts them
  23 // if possible.
  24 //
  25 // A Scheduler unlocks any containers that are locked but can't be
  26 // mapped. (For example, this happens when the cloud provider reaches
  27 // quota/capacity and a previously mappable container's priority is
  28 // surpassed by a newer container.)
  29 //
  30 // If it encounters errors while creating new workers, a Scheduler
  31 // shuts down idle workers, in case they are consuming quota.
  32 type Scheduler struct {
  33         logger              logrus.FieldLogger
  34         client              *arvados.Client
  35         queue               ContainerQueue
  36         pool                WorkerPool
  37         reg                 *prometheus.Registry
  38         staleLockTimeout    time.Duration
  39         queueUpdateInterval time.Duration
  40
  41         uuidOp map[string]string // operation in progress: "lock", "cancel", ...
  42         mtx    sync.Mutex
  43         wakeup *time.Timer
  44
  45         runOnce sync.Once
  46         stop    chan struct{}
  47         stopped chan struct{}
  48
  49         last503time    time.Time // last time API responded 503
  50         maxConcurrency int       // dynamic container limit (0 = unlimited), see runQueue()
  51
  52         mContainersAllocatedNotStarted   prometheus.Gauge
  53         mContainersNotAllocatedOverQuota prometheus.Gauge
  54         mLongestWaitTimeSinceQueue       prometheus.Gauge
  55 }
  56
  57 // New returns a new unstarted Scheduler.
  58 //
  59 // Any given queue and pool should not be used by more than one
  60 // scheduler at a time.
  61 func New(ctx context.Context, client *arvados.Client, queue ContainerQueue, pool WorkerPool, reg *prometheus.Registry, staleLockTimeout, queueUpdateInterval time.Duration) *Scheduler {
  62         sch := &Scheduler{
  63                 logger:              ctxlog.FromContext(ctx),
  64                 client:              client,
  65                 queue:               queue,
  66                 pool:                pool,
  67                 reg:                 reg,
  68                 staleLockTimeout:    staleLockTimeout,
  69                 queueUpdateInterval: queueUpdateInterval,
  70                 wakeup:              time.NewTimer(time.Second),
  71                 stop:                make(chan struct{}),
  72                 stopped:             make(chan struct{}),
  73                 uuidOp:              map[string]string{},
  74         }
  75         sch.registerMetrics(reg)
  76         return sch
  77 }
  78
  79 func (sch *Scheduler) registerMetrics(reg *prometheus.Registry) {
  80         if reg == nil {
  81                 reg = prometheus.NewRegistry()
  82         }
  83         sch.mContainersAllocatedNotStarted = prometheus.NewGauge(prometheus.GaugeOpts{
  84                 Namespace: "arvados",
  85                 Subsystem: "dispatchcloud",
  86                 Name:      "containers_allocated_not_started",
  87                 Help:      "Number of containers allocated to a worker but not started yet (worker is booting).",
  88         })
  89         reg.MustRegister(sch.mContainersAllocatedNotStarted)
  90         sch.mContainersNotAllocatedOverQuota = prometheus.NewGauge(prometheus.GaugeOpts{
  91                 Namespace: "arvados",
  92                 Subsystem: "dispatchcloud",
  93                 Name:      "containers_not_allocated_over_quota",
  94                 Help:      "Number of containers not allocated to a worker because the system has hit a quota.",
  95         })
  96         reg.MustRegister(sch.mContainersNotAllocatedOverQuota)
  97         sch.mLongestWaitTimeSinceQueue = prometheus.NewGauge(prometheus.GaugeOpts{
  98                 Namespace: "arvados",
  99                 Subsystem: "dispatchcloud",
 100                 Name:      "containers_longest_wait_time_seconds",
 101                 Help:      "Current longest wait time of any container since queuing, and before the start of crunch-run.",
 102         })
 103         reg.MustRegister(sch.mLongestWaitTimeSinceQueue)
 104 }
 105
 106 func (sch *Scheduler) updateMetrics() {
 107         earliest := time.Time{}
 108         entries, _ := sch.queue.Entries()
 109         running := sch.pool.Running()
 110         for _, ent := range entries {
 111                 if ent.Container.Priority > 0 &&
 112                         (ent.Container.State == arvados.ContainerStateQueued || ent.Container.State == arvados.ContainerStateLocked) {
 113                         // Exclude containers that are preparing to run the payload (i.e.
 114                         // ContainerStateLocked and running on a worker, most likely loading the
 115                         // payload image
 116                         if _, ok := running[ent.Container.UUID]; !ok {
 117                                 if ent.Container.CreatedAt.Before(earliest) || earliest.IsZero() {
 118                                         earliest = ent.Container.CreatedAt
 119                                 }
 120                         }
 121                 }
 122         }
 123         if !earliest.IsZero() {
 124                 sch.mLongestWaitTimeSinceQueue.Set(time.Since(earliest).Seconds())
 125         } else {
 126                 sch.mLongestWaitTimeSinceQueue.Set(0)
 127         }
 128 }
 129
 130 // Start starts the scheduler.
 131 func (sch *Scheduler) Start() {
 132         go sch.runOnce.Do(sch.run)
 133 }
 134
 135 // Stop stops the scheduler. No other method should be called after
 136 // Stop.
 137 func (sch *Scheduler) Stop() {
 138         close(sch.stop)
 139         <-sch.stopped
 140 }
 141
 142 func (sch *Scheduler) run() {
 143         defer close(sch.stopped)
 144
 145         // Ensure the queue is fetched once before attempting anything.
 146         for err := sch.queue.Update(); err != nil; err = sch.queue.Update() {
 147                 sch.logger.Errorf("error updating queue: %s", err)
 148                 d := sch.queueUpdateInterval / 10
 149                 if d < time.Second {
 150                         d = time.Second
 151                 }
 152                 sch.logger.Infof("waiting %s before retry", d)
 153                 time.Sleep(d)
 154         }
 155
 156         // Keep the queue up to date.
 157         poll := time.NewTicker(sch.queueUpdateInterval)
 158         defer poll.Stop()
 159         go func() {
 160                 for range poll.C {
 161                         err := sch.queue.Update()
 162                         if err != nil {
 163                                 sch.logger.Errorf("error updating queue: %s", err)
 164                         }
 165                 }
 166         }()
 167
 168         t0 := time.Now()
 169         sch.logger.Infof("FixStaleLocks starting.")
 170         sch.fixStaleLocks()
 171         sch.logger.Infof("FixStaleLocks finished (%s), starting scheduling.", time.Since(t0))
 172
 173         poolNotify := sch.pool.Subscribe()
 174         defer sch.pool.Unsubscribe(poolNotify)
 175
 176         queueNotify := sch.queue.Subscribe()
 177         defer sch.queue.Unsubscribe(queueNotify)
 178
 179         for {
 180                 sch.runQueue()
 181                 sch.sync()
 182                 sch.updateMetrics()
 183                 select {
 184                 case <-sch.stop:
 185                         return
 186                 case <-queueNotify:
 187                 case <-poolNotify:
 188                 case <-sch.wakeup.C:
 189                 }
 190         }
 191 }