// populated.
Container arvados.Container `json:"container"`
InstanceType arvados.InstanceType `json:"instance_type"`
+ FirstSeenAt time.Time `json:"first_seen_at"`
}
// String implements fmt.Stringer by returning the queued container's
delete(cq.current, uuid)
}
+// Caller must have lock.
func (cq *Queue) addEnt(uuid string, ctr arvados.Container) {
it, err := cq.chooseType(&ctr)
if err != nil && (ctr.State == arvados.ContainerStateQueued || ctr.State == arvados.ContainerStateLocked) {
"Priority": ctr.Priority,
"InstanceType": it.Name,
}).Info("adding container to queue")
- cq.current[uuid] = QueueEnt{Container: ctr, InstanceType: it}
+ cq.current[uuid] = QueueEnt{Container: ctr, InstanceType: it, FirstSeenAt: time.Now()}
}
// Lock acquires the dispatch lock for the given container.
sorted = append(sorted, ent)
}
sort.Slice(sorted, func(i, j int) bool {
- return sorted[i].Container.Priority > sorted[j].Container.Priority
+ if pi, pj := sorted[i].Container.Priority, sorted[j].Container.Priority; pi != pj {
+ return pi > pj
+ } else {
+ // When containers have identical priority,
+ // start them in the order we first noticed
+ // them. This avoids extra lock/unlock cycles
+ // when we unlock the containers that don't
+ // fit in the available pool.
+ return sorted[i].FirstSeenAt.Before(sorted[j].FirstSeenAt)
+ }
})
running := sch.pool.Running()
}
}
+// Don't flap lock/unlock when equal-priority containers compete for
+// limited workers.
+//
+// (Unless we use FirstSeenAt as a secondary sort key, each runQueue()
+// tends to choose a different one of the equal-priority containers as
+// the "first" one that should be locked, and unlock the one it chose
+// last time. This generates logging noise, and fails containers by
+// reaching MaxDispatchAttempts quickly.)
+func (*SchedulerSuite) TestEqualPriorityContainers(c *check.C) {
+ logger := ctxlog.TestLogger(c)
+ ctx := ctxlog.Context(context.Background(), logger)
+ queue := test.Queue{
+ ChooseType: chooseType,
+ Logger: logger,
+ }
+ for i := 0; i < 8; i++ {
+ queue.Containers = append(queue.Containers, arvados.Container{
+ UUID: test.ContainerUUID(i),
+ Priority: 333,
+ State: arvados.ContainerStateQueued,
+ RuntimeConstraints: arvados.RuntimeConstraints{
+ VCPUs: 3,
+ RAM: 3 << 30,
+ },
+ })
+ }
+ queue.Update()
+ pool := stubPool{
+ quota: 2,
+ unalloc: map[arvados.InstanceType]int{
+ test.InstanceType(3): 1,
+ },
+ idle: map[arvados.InstanceType]int{
+ test.InstanceType(3): 1,
+ },
+ running: map[string]time.Time{},
+ creates: []arvados.InstanceType{},
+ starts: []string{},
+ canCreate: 1,
+ }
+ sch := New(ctx, &queue, &pool, nil, time.Millisecond, time.Millisecond)
+ for i := 0; i < 30; i++ {
+ sch.runQueue()
+ sch.sync()
+ time.Sleep(time.Millisecond)
+ }
+ c.Check(pool.shutdowns, check.Equals, 0)
+ c.Check(pool.starts, check.HasLen, 1)
+ unlocked := map[string]int{}
+ for _, chg := range queue.StateChanges() {
+ if chg.To == arvados.ContainerStateQueued {
+ unlocked[chg.UUID]++
+ }
+ }
+ for uuid, count := range unlocked {
+ c.Check(count, check.Equals, 1, check.Commentf("%s", uuid))
+ }
+}
+
// Start lower-priority containers while waiting for new/existing
// workers to come up for higher-priority containers.
func (*SchedulerSuite) TestStartWhileCreating(c *check.C) {