1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
7 // Dispatcher service for Crunch that submits containers to the slurm queue.
20 "git.curoverse.com/arvados.git/lib/dispatchcloud"
21 "git.curoverse.com/arvados.git/sdk/go/arvados"
22 "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
23 "git.curoverse.com/arvados.git/sdk/go/config"
24 "git.curoverse.com/arvados.git/sdk/go/dispatch"
25 "github.com/coreos/go-systemd/daemon"
28 const initialNiceValue int64 = 10000
32 defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
35 type Dispatcher struct {
37 cluster *arvados.Cluster
38 sqCheck *SqueueChecker
43 SbatchArguments []string
44 PollPeriod arvados.Duration
47 // crunch-run command to invoke. The container UUID will be
48 // appended. If nil, []string{"crunch-run"} will be used.
50 // Example: []string{"crunch-run", "--cgroup-parent-subsystem=memory"}
51 CrunchRunCommand []string
53 // Extra RAM to reserve (in Bytes) for SLURM job, in addition
54 // to the amount specified in the container's RuntimeConstraints
57 // Minimum time between two attempts to run the same container
58 MinRetryPeriod arvados.Duration
63 err := disp.Run(os.Args[0], os.Args[1:])
69 func (disp *Dispatcher) Run(prog string, args []string) error {
70 if err := disp.configure(prog, args); err != nil {
77 // configure() loads config files. Tests skip this.
78 func (disp *Dispatcher) configure(prog string, args []string) error {
79 flags := flag.NewFlagSet(prog, flag.ExitOnError)
80 flags.Usage = func() { usage(flags) }
82 configPath := flags.String(
85 "`path` to JSON or YAML configuration file")
86 dumpConfig := flag.Bool(
89 "write current configuration to stdout and exit")
90 getVersion := flags.Bool(
93 "Print version information and exit.")
94 // Parse args; omit the first arg which is the command name
97 // Print version information if requested
99 fmt.Printf("crunch-dispatch-slurm %s\n", version)
103 log.Printf("crunch-dispatch-slurm %s started", version)
105 err := disp.readConfig(*configPath)
110 if disp.CrunchRunCommand == nil {
111 disp.CrunchRunCommand = []string{"crunch-run"}
114 if disp.PollPeriod == 0 {
115 disp.PollPeriod = arvados.Duration(10 * time.Second)
118 if disp.Client.APIHost != "" || disp.Client.AuthToken != "" {
119 // Copy real configs into env vars so [a]
120 // MakeArvadosClient() uses them, and [b] they get
121 // propagated to crunch-run via SLURM.
122 os.Setenv("ARVADOS_API_HOST", disp.Client.APIHost)
123 os.Setenv("ARVADOS_API_TOKEN", disp.Client.AuthToken)
124 os.Setenv("ARVADOS_API_HOST_INSECURE", "")
125 if disp.Client.Insecure {
126 os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
128 os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(disp.Client.KeepServiceURIs, " "))
129 os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
131 log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).")
135 return config.DumpAndExit(disp)
138 siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile)
139 if os.IsNotExist(err) {
140 log.Printf("warning: no cluster config (%s), proceeding with no node types defined", err)
141 } else if err != nil {
142 return fmt.Errorf("error loading config: %s", err)
143 } else if disp.cluster, err = siteConfig.GetCluster(""); err != nil {
144 return fmt.Errorf("config error: %s", err)
150 // setup() initializes private fields after configure().
151 func (disp *Dispatcher) setup() {
152 arv, err := arvadosclient.MakeArvadosClient()
154 log.Fatalf("Error making Arvados client: %v", err)
158 disp.slurm = &slurmCLI{}
159 disp.sqCheck = &SqueueChecker{
160 Period: time.Duration(disp.PollPeriod),
161 PrioritySpread: disp.PrioritySpread,
164 disp.Dispatcher = &dispatch.Dispatcher{
166 RunContainer: disp.runContainer,
167 PollPeriod: time.Duration(disp.PollPeriod),
168 MinRetryPeriod: time.Duration(disp.MinRetryPeriod),
172 func (disp *Dispatcher) run() error {
173 defer disp.sqCheck.Stop()
175 if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 {
176 go dispatchcloud.SlurmNodeTypeFeatureKludge(disp.cluster)
179 if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
180 log.Printf("Error notifying init daemon: %v", err)
182 go disp.checkSqueueForOrphans()
183 return disp.Dispatcher.Run(context.Background())
186 var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$`)
188 // Check the next squeue report, and invoke TrackContainer for all the
189 // containers in the report. This gives us a chance to cancel slurm
190 // jobs started by a previous dispatch process that never released
191 // their slurm allocations even though their container states are
192 // Cancelled or Complete. See https://dev.arvados.org/issues/10979
193 func (disp *Dispatcher) checkSqueueForOrphans() {
194 for _, uuid := range disp.sqCheck.All() {
195 if !containerUuidPattern.MatchString(uuid) {
198 err := disp.TrackContainer(uuid)
200 log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err)
205 func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) {
206 mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM+disp.ReserveExtraRAM) / float64(1048576)))
209 for _, m := range container.Mounts {
214 disk = int64(math.Ceil(float64(disk) / float64(1048576)))
216 var sbatchArgs []string
217 sbatchArgs = append(sbatchArgs, disp.SbatchArguments...)
218 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--job-name=%s", container.UUID))
219 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--mem=%d", mem))
220 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs))
221 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--tmp=%d", disk))
222 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--nice=%d", initialNiceValue))
223 if len(container.SchedulingParameters.Partitions) > 0 {
224 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ",")))
227 if disp.cluster == nil {
228 // no instance types configured
229 } else if it, err := dispatchcloud.ChooseInstanceType(disp.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured {
231 } else if err != nil {
234 sbatchArgs = append(sbatchArgs, "--constraint=instancetype="+it.Name)
237 return sbatchArgs, nil
240 func (disp *Dispatcher) submit(container arvados.Container, crunchRunCommand []string) error {
241 // append() here avoids modifying crunchRunCommand's
242 // underlying array, which is shared with other goroutines.
243 crArgs := append([]string(nil), crunchRunCommand...)
244 crArgs = append(crArgs, container.UUID)
245 crScript := strings.NewReader(execScript(crArgs))
247 disp.sqCheck.L.Lock()
248 defer disp.sqCheck.L.Unlock()
250 sbArgs, err := disp.sbatchArgs(container)
254 log.Printf("running sbatch %+q", sbArgs)
255 return disp.slurm.Batch(crScript, sbArgs)
258 // Submit a container to the slurm queue (or resume monitoring if it's
259 // already in the queue). Cancel the slurm job if the container's
260 // priority changes to zero or its state indicates it's no longer
262 func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
263 ctx, cancel := context.WithCancel(context.Background())
266 if ctr.State == dispatch.Locked && !disp.sqCheck.HasUUID(ctr.UUID) {
267 log.Printf("Submitting container %s to slurm", ctr.UUID)
268 if err := disp.submit(ctr, disp.CrunchRunCommand); err != nil {
270 if err == dispatchcloud.ErrConstraintsNotSatisfiable {
271 text = fmt.Sprintf("cannot run container %s: %s", ctr.UUID, err)
272 disp.UpdateState(ctr.UUID, dispatch.Cancelled)
274 text = fmt.Sprintf("Error submitting container %s to slurm: %s", ctr.UUID, err)
278 lr := arvadosclient.Dict{"log": arvadosclient.Dict{
279 "object_uuid": ctr.UUID,
280 "event_type": "dispatch",
281 "properties": map[string]string{"text": text}}}
282 disp.Arv.Create("logs", lr, nil)
284 disp.Unlock(ctr.UUID)
289 log.Printf("Start monitoring container %v in state %q", ctr.UUID, ctr.State)
290 defer log.Printf("Done monitoring container %s", ctr.UUID)
292 // If the container disappears from the slurm queue, there is
293 // no point in waiting for further dispatch updates: just
294 // clean up and return.
295 go func(uuid string) {
296 for ctx.Err() == nil && disp.sqCheck.HasUUID(uuid) {
304 // Disappeared from squeue
305 if err := disp.Arv.Get("containers", ctr.UUID, nil, &ctr); err != nil {
306 log.Printf("Error getting final container state for %s: %s", ctr.UUID, err)
309 case dispatch.Running:
310 disp.UpdateState(ctr.UUID, dispatch.Cancelled)
311 case dispatch.Locked:
312 disp.Unlock(ctr.UUID)
315 case updated, ok := <-status:
317 log.Printf("container %s is done: cancel slurm job", ctr.UUID)
319 } else if updated.Priority == 0 {
320 log.Printf("container %s has state %q, priority %d: cancel slurm job", ctr.UUID, updated.State, updated.Priority)
323 p := int64(updated.Priority)
326 // user-assigned priority. If
327 // ctrs have equal priority,
328 // run the older one first.
329 p = int64(p)<<50 - (updated.CreatedAt.UnixNano() >> 14)
331 disp.sqCheck.SetPriority(ctr.UUID, p)
336 func (disp *Dispatcher) scancel(ctr arvados.Container) {
337 disp.sqCheck.L.Lock()
338 err := disp.slurm.Cancel(ctr.UUID)
339 disp.sqCheck.L.Unlock()
342 log.Printf("scancel: %s", err)
343 time.Sleep(time.Second)
344 } else if disp.sqCheck.HasUUID(ctr.UUID) {
345 log.Printf("container %s is still in squeue after scancel", ctr.UUID)
346 time.Sleep(time.Second)
350 func (disp *Dispatcher) readConfig(path string) error {
351 err := config.LoadFile(disp, path)
352 if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
353 log.Printf("Config not specified. Continue with default configuration.")