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.
21 "git.curoverse.com/arvados.git/lib/dispatchcloud"
22 "git.curoverse.com/arvados.git/sdk/go/arvados"
23 "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
24 "git.curoverse.com/arvados.git/sdk/go/config"
25 "git.curoverse.com/arvados.git/sdk/go/dispatch"
26 "github.com/Sirupsen/logrus"
27 "github.com/coreos/go-systemd/daemon"
30 type logger interface {
32 Fatalf(string, ...interface{})
35 const initialNiceValue int64 = 10000
39 defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
42 type Dispatcher struct {
44 logger logrus.FieldLogger
45 cluster *arvados.Cluster
46 sqCheck *SqueueChecker
51 SbatchArguments []string
52 PollPeriod arvados.Duration
55 // crunch-run command to invoke. The container UUID will be
56 // appended. If nil, []string{"crunch-run"} will be used.
58 // Example: []string{"crunch-run", "--cgroup-parent-subsystem=memory"}
59 CrunchRunCommand []string
61 // Extra RAM to reserve (in Bytes) for SLURM job, in addition
62 // to the amount specified in the container's RuntimeConstraints
65 // Minimum time between two attempts to run the same container
66 MinRetryPeriod arvados.Duration
70 logger := logrus.StandardLogger()
71 if os.Getenv("DEBUG") != "" {
72 logger.SetLevel(logrus.DebugLevel)
74 logger.Formatter = &logrus.JSONFormatter{
75 TimestampFormat: "2006-01-02T15:04:05.000000000Z07:00",
77 disp := &Dispatcher{logger: logger}
78 err := disp.Run(os.Args[0], os.Args[1:])
80 logrus.Fatalf("%s", err)
84 func (disp *Dispatcher) Run(prog string, args []string) error {
85 if err := disp.configure(prog, args); err != nil {
92 // configure() loads config files. Tests skip this.
93 func (disp *Dispatcher) configure(prog string, args []string) error {
94 flags := flag.NewFlagSet(prog, flag.ExitOnError)
95 flags.Usage = func() { usage(flags) }
97 configPath := flags.String(
100 "`path` to JSON or YAML configuration file")
101 dumpConfig := flag.Bool(
104 "write current configuration to stdout and exit")
105 getVersion := flags.Bool(
108 "Print version information and exit.")
109 // Parse args; omit the first arg which is the command name
112 // Print version information if requested
114 fmt.Printf("crunch-dispatch-slurm %s\n", version)
118 disp.logger.Printf("crunch-dispatch-slurm %s started", version)
120 err := disp.readConfig(*configPath)
125 if disp.CrunchRunCommand == nil {
126 disp.CrunchRunCommand = []string{"crunch-run"}
129 if disp.PollPeriod == 0 {
130 disp.PollPeriod = arvados.Duration(10 * time.Second)
133 if disp.Client.APIHost != "" || disp.Client.AuthToken != "" {
134 // Copy real configs into env vars so [a]
135 // MakeArvadosClient() uses them, and [b] they get
136 // propagated to crunch-run via SLURM.
137 os.Setenv("ARVADOS_API_HOST", disp.Client.APIHost)
138 os.Setenv("ARVADOS_API_TOKEN", disp.Client.AuthToken)
139 os.Setenv("ARVADOS_API_HOST_INSECURE", "")
140 if disp.Client.Insecure {
141 os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
143 os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(disp.Client.KeepServiceURIs, " "))
144 os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
146 disp.logger.Warnf("Client credentials missing from config, so falling back on environment variables (deprecated).")
150 return config.DumpAndExit(disp)
153 siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile)
154 if os.IsNotExist(err) {
155 disp.logger.Warnf("no cluster config (%s), proceeding with no node types defined", err)
156 } else if err != nil {
157 return fmt.Errorf("error loading config: %s", err)
158 } else if disp.cluster, err = siteConfig.GetCluster(""); err != nil {
159 return fmt.Errorf("config error: %s", err)
165 // setup() initializes private fields after configure().
166 func (disp *Dispatcher) setup() {
167 if disp.logger == nil {
168 disp.logger = logrus.StandardLogger()
170 arv, err := arvadosclient.MakeArvadosClient()
172 disp.logger.Fatalf("Error making Arvados client: %v", err)
176 disp.slurm = &slurmCLI{}
177 disp.sqCheck = &SqueueChecker{
179 Period: time.Duration(disp.PollPeriod),
180 PrioritySpread: disp.PrioritySpread,
183 disp.Dispatcher = &dispatch.Dispatcher{
186 RunContainer: disp.runContainer,
187 PollPeriod: time.Duration(disp.PollPeriod),
188 MinRetryPeriod: time.Duration(disp.MinRetryPeriod),
192 func (disp *Dispatcher) run() error {
193 defer disp.sqCheck.Stop()
195 if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 {
196 go dispatchcloud.SlurmNodeTypeFeatureKludge(disp.cluster)
199 if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
200 log.Printf("Error notifying init daemon: %v", err)
202 go disp.checkSqueueForOrphans()
203 return disp.Dispatcher.Run(context.Background())
206 var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$`)
208 // Check the next squeue report, and invoke TrackContainer for all the
209 // containers in the report. This gives us a chance to cancel slurm
210 // jobs started by a previous dispatch process that never released
211 // their slurm allocations even though their container states are
212 // Cancelled or Complete. See https://dev.arvados.org/issues/10979
213 func (disp *Dispatcher) checkSqueueForOrphans() {
214 for _, uuid := range disp.sqCheck.All() {
215 if !containerUuidPattern.MatchString(uuid) {
218 err := disp.TrackContainer(uuid)
220 log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err)
225 func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []string {
226 mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM+disp.ReserveExtraRAM) / float64(1048576)))
229 for _, m := range container.Mounts {
234 disk = int64(math.Ceil(float64(disk) / float64(1048576)))
236 fmt.Sprintf("--mem=%d", mem),
237 fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs),
238 fmt.Sprintf("--tmp=%d", disk),
242 func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) {
244 args = append(args, disp.SbatchArguments...)
245 args = append(args, "--job-name="+container.UUID, fmt.Sprintf("--nice=%d", initialNiceValue))
247 if disp.cluster == nil {
248 // no instance types configured
249 args = append(args, disp.slurmConstraintArgs(container)...)
250 } else if it, err := dispatchcloud.ChooseInstanceType(disp.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured {
252 args = append(args, disp.slurmConstraintArgs(container)...)
253 } else if err != nil {
256 // use instancetype constraint instead of slurm mem/cpu/tmp specs
257 args = append(args, "--constraint=instancetype="+it.Name)
260 if len(container.SchedulingParameters.Partitions) > 0 {
261 args = append(args, "--partition="+strings.Join(container.SchedulingParameters.Partitions, ","))
267 func (disp *Dispatcher) submit(container arvados.Container, crunchRunCommand []string) error {
268 // append() here avoids modifying crunchRunCommand's
269 // underlying array, which is shared with other goroutines.
270 crArgs := append([]string(nil), crunchRunCommand...)
271 crArgs = append(crArgs, container.UUID)
272 crScript := strings.NewReader(execScript(crArgs))
274 sbArgs, err := disp.sbatchArgs(container)
278 log.Printf("running sbatch %+q", sbArgs)
279 return disp.slurm.Batch(crScript, sbArgs)
282 // Submit a container to the slurm queue (or resume monitoring if it's
283 // already in the queue). Cancel the slurm job if the container's
284 // priority changes to zero or its state indicates it's no longer
286 func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
287 ctx, cancel := context.WithCancel(context.Background())
290 if ctr.State == dispatch.Locked && !disp.sqCheck.HasUUID(ctr.UUID) {
291 log.Printf("Submitting container %s to slurm", ctr.UUID)
292 if err := disp.submit(ctr, disp.CrunchRunCommand); err != nil {
294 if err, ok := err.(dispatchcloud.ConstraintsNotSatisfiableError); ok {
295 var logBuf bytes.Buffer
296 fmt.Fprintf(&logBuf, "cannot run container %s: %s\n", ctr.UUID, err)
297 if len(err.AvailableTypes) == 0 {
298 fmt.Fprint(&logBuf, "No instance types are configured.\n")
300 fmt.Fprint(&logBuf, "Available instance types:\n")
301 for _, t := range err.AvailableTypes {
303 "Type %q: %d VCPUs, %d RAM, %d Scratch, %f Price\n",
304 t.Name, t.VCPUs, t.RAM, t.Scratch, t.Price,
308 text = logBuf.String()
309 disp.UpdateState(ctr.UUID, dispatch.Cancelled)
311 text = fmt.Sprintf("Error submitting container %s to slurm: %s", ctr.UUID, err)
315 lr := arvadosclient.Dict{"log": arvadosclient.Dict{
316 "object_uuid": ctr.UUID,
317 "event_type": "dispatch",
318 "properties": map[string]string{"text": text}}}
319 disp.Arv.Create("logs", lr, nil)
321 disp.Unlock(ctr.UUID)
326 log.Printf("Start monitoring container %v in state %q", ctr.UUID, ctr.State)
327 defer log.Printf("Done monitoring container %s", ctr.UUID)
329 // If the container disappears from the slurm queue, there is
330 // no point in waiting for further dispatch updates: just
331 // clean up and return.
332 go func(uuid string) {
333 for ctx.Err() == nil && disp.sqCheck.HasUUID(uuid) {
341 // Disappeared from squeue
342 if err := disp.Arv.Get("containers", ctr.UUID, nil, &ctr); err != nil {
343 log.Printf("error getting final container state for %s: %s", ctr.UUID, err)
346 case dispatch.Running:
347 disp.UpdateState(ctr.UUID, dispatch.Cancelled)
348 case dispatch.Locked:
349 disp.Unlock(ctr.UUID)
352 case updated, ok := <-status:
354 log.Printf("container %s is done: cancel slurm job", ctr.UUID)
356 } else if updated.Priority == 0 {
357 log.Printf("container %s has state %q, priority %d: cancel slurm job", ctr.UUID, updated.State, updated.Priority)
360 p := int64(updated.Priority)
363 // user-assigned priority. If
364 // ctrs have equal priority,
365 // run the older one first.
366 p = int64(p)<<50 - (updated.CreatedAt.UnixNano() >> 14)
368 disp.sqCheck.SetPriority(ctr.UUID, p)
373 func (disp *Dispatcher) scancel(ctr arvados.Container) {
374 err := disp.slurm.Cancel(ctr.UUID)
376 log.Printf("scancel: %s", err)
377 time.Sleep(time.Second)
378 } else if disp.sqCheck.HasUUID(ctr.UUID) {
379 log.Printf("container %s is still in squeue after scancel", ctr.UUID)
380 time.Sleep(time.Second)
384 func (disp *Dispatcher) readConfig(path string) error {
385 err := config.LoadFile(disp, path)
386 if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
387 log.Printf("Config not specified. Continue with default configuration.")