1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
17 type slurmJob struct {
20 priority int64 // current slurm priority (incorporates nice value)
21 nice int64 // current slurm nice value
24 // Squeue implements asynchronous polling monitor of the SLURM queue using the
26 type SqueueChecker struct {
30 queue map[string]*slurmJob
36 // HasUUID checks if a given container UUID is in the slurm queue.
37 // This does not run squeue directly, but instead blocks until woken
38 // up by next successful update of squeue.
39 func (sqc *SqueueChecker) HasUUID(uuid string) bool {
40 sqc.startOnce.Do(sqc.start)
45 // block until next squeue broadcast signaling an update.
47 _, exists := sqc.queue[uuid]
51 // SetPriority sets or updates the desired (Arvados) priority for a
53 func (sqc *SqueueChecker) SetPriority(uuid string, want int64) {
54 sqc.startOnce.Do(sqc.start)
57 job, ok := sqc.queue[uuid]
59 // Wait in case the slurm job was just submitted and
60 // will appear in the next squeue update.
62 if job, ok = sqc.queue[uuid]; !ok {
66 job.wantPriority = want
69 // adjust slurm job nice values as needed to ensure slurm priority
70 // order matches Arvados priority order.
71 func (sqc *SqueueChecker) reniceAll() {
75 jobs := make([]*slurmJob, 0, len(sqc.queue))
76 for _, j := range sqc.queue {
77 if j.wantPriority == 0 {
78 // SLURM job with unknown Arvados priority
79 // (perhaps it's not an Arvados job)
82 jobs = append(jobs, j)
85 sort.Slice(jobs, func(i, j int) bool {
86 return jobs[i].wantPriority > jobs[j].wantPriority
88 renice := wantNice(jobs, sqc.PrioritySpread)
89 for i, job := range jobs {
90 if renice[i] == job.nice {
93 log.Printf("updating slurm priority for %q: nice %d => %d", job.uuid, job.nice, renice[i])
94 sqc.Slurm.Renice(job.uuid, renice[i])
98 // Stop stops the squeue monitoring goroutine. Do not call HasUUID
99 // after calling Stop.
100 func (sqc *SqueueChecker) Stop() {
106 // check gets the names of jobs in the SLURM queue (running and
107 // queued). If it succeeds, it updates sqc.queue and wakes up any
108 // goroutines that are waiting in HasUUID() or All().
109 func (sqc *SqueueChecker) check() {
110 // Mutex between squeue sync and running sbatch or scancel. This
111 // establishes a sequence so that squeue doesn't run concurrently with
112 // sbatch or scancel; the next update of squeue will occur only after
113 // sbatch or scancel has completed.
117 cmd := sqc.Slurm.QueueCommand([]string{"--all", "--format=%j %y %Q"})
118 stdout, stderr := &bytes.Buffer{}, &bytes.Buffer{}
119 cmd.Stdout, cmd.Stderr = stdout, stderr
120 if err := cmd.Run(); err != nil {
121 log.Printf("Error running %q %q: %s %q", cmd.Path, cmd.Args, err, stderr.String())
125 lines := strings.Split(stdout.String(), "\n")
126 newq := make(map[string]*slurmJob, len(lines))
127 for _, line := range lines {
133 if _, err := fmt.Sscan(line, &uuid, &n, &p); err != nil {
134 log.Printf("warning: ignoring unparsed line in squeue output: %q", line)
137 replacing, ok := sqc.queue[uuid]
139 replacing = &slurmJob{uuid: uuid}
141 replacing.priority = p
143 newq[uuid] = replacing
149 // Initialize, and start a goroutine to call check() once per
150 // squeue.Period until terminated by calling Stop().
151 func (sqc *SqueueChecker) start() {
152 sqc.L = &sync.Mutex{}
153 sqc.done = make(chan struct{})
155 ticker := time.NewTicker(sqc.Period)
169 // All waits for the next squeue invocation, and returns all job
170 // names reported by squeue.
171 func (sqc *SqueueChecker) All() []string {
172 sqc.startOnce.Do(sqc.start)
177 for u := range sqc.queue {
178 uuids = append(uuids, u)