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
68 // Batch size for container queries
73 logger := logrus.StandardLogger()
74 if os.Getenv("DEBUG") != "" {
75 logger.SetLevel(logrus.DebugLevel)
77 logger.Formatter = &logrus.JSONFormatter{
78 TimestampFormat: "2006-01-02T15:04:05.000000000Z07:00",
80 disp := &Dispatcher{logger: logger}
81 err := disp.Run(os.Args[0], os.Args[1:])
83 logrus.Fatalf("%s", err)
87 func (disp *Dispatcher) Run(prog string, args []string) error {
88 if err := disp.configure(prog, args); err != nil {
95 // configure() loads config files. Tests skip this.
96 func (disp *Dispatcher) configure(prog string, args []string) error {
97 flags := flag.NewFlagSet(prog, flag.ExitOnError)
98 flags.Usage = func() { usage(flags) }
100 configPath := flags.String(
103 "`path` to JSON or YAML configuration file")
104 dumpConfig := flag.Bool(
107 "write current configuration to stdout and exit")
108 getVersion := flags.Bool(
111 "Print version information and exit.")
112 // Parse args; omit the first arg which is the command name
115 // Print version information if requested
117 fmt.Printf("crunch-dispatch-slurm %s\n", version)
121 disp.logger.Printf("crunch-dispatch-slurm %s started", version)
123 err := disp.readConfig(*configPath)
128 if disp.CrunchRunCommand == nil {
129 disp.CrunchRunCommand = []string{"crunch-run"}
132 if disp.PollPeriod == 0 {
133 disp.PollPeriod = arvados.Duration(10 * time.Second)
136 if disp.Client.APIHost != "" || disp.Client.AuthToken != "" {
137 // Copy real configs into env vars so [a]
138 // MakeArvadosClient() uses them, and [b] they get
139 // propagated to crunch-run via SLURM.
140 os.Setenv("ARVADOS_API_HOST", disp.Client.APIHost)
141 os.Setenv("ARVADOS_API_TOKEN", disp.Client.AuthToken)
142 os.Setenv("ARVADOS_API_HOST_INSECURE", "")
143 if disp.Client.Insecure {
144 os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
146 os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(disp.Client.KeepServiceURIs, " "))
147 os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
149 disp.logger.Warnf("Client credentials missing from config, so falling back on environment variables (deprecated).")
153 return config.DumpAndExit(disp)
156 siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile)
157 if os.IsNotExist(err) {
158 disp.logger.Warnf("no cluster config (%s), proceeding with no node types defined", err)
159 } else if err != nil {
160 return fmt.Errorf("error loading config: %s", err)
161 } else if disp.cluster, err = siteConfig.GetCluster(""); err != nil {
162 return fmt.Errorf("config error: %s", err)
168 // setup() initializes private fields after configure().
169 func (disp *Dispatcher) setup() {
170 if disp.logger == nil {
171 disp.logger = logrus.StandardLogger()
173 arv, err := arvadosclient.MakeArvadosClient()
175 disp.logger.Fatalf("Error making Arvados client: %v", err)
179 disp.slurm = NewSlurmCLI()
180 disp.sqCheck = &SqueueChecker{
182 Period: time.Duration(disp.PollPeriod),
183 PrioritySpread: disp.PrioritySpread,
186 disp.Dispatcher = &dispatch.Dispatcher{
189 BatchSize: disp.BatchSize,
190 RunContainer: disp.runContainer,
191 PollPeriod: time.Duration(disp.PollPeriod),
192 MinRetryPeriod: time.Duration(disp.MinRetryPeriod),
196 func (disp *Dispatcher) run() error {
197 defer disp.sqCheck.Stop()
199 if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 {
200 go SlurmNodeTypeFeatureKludge(disp.cluster)
203 if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
204 log.Printf("Error notifying init daemon: %v", err)
206 go disp.checkSqueueForOrphans()
207 return disp.Dispatcher.Run(context.Background())
210 var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$`)
212 // Check the next squeue report, and invoke TrackContainer for all the
213 // containers in the report. This gives us a chance to cancel slurm
214 // jobs started by a previous dispatch process that never released
215 // their slurm allocations even though their container states are
216 // Cancelled or Complete. See https://dev.arvados.org/issues/10979
217 func (disp *Dispatcher) checkSqueueForOrphans() {
218 for _, uuid := range disp.sqCheck.All() {
219 if !containerUuidPattern.MatchString(uuid) {
222 err := disp.TrackContainer(uuid)
224 log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err)
229 func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []string {
230 mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM+disp.ReserveExtraRAM) / float64(1048576)))
232 disk := dispatchcloud.EstimateScratchSpace(&container)
233 disk = int64(math.Ceil(float64(disk) / float64(1048576)))
235 fmt.Sprintf("--mem=%d", mem),
236 fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs),
237 fmt.Sprintf("--tmp=%d", disk),
241 func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) {
243 args = append(args, disp.SbatchArguments...)
244 args = append(args, "--job-name="+container.UUID, fmt.Sprintf("--nice=%d", initialNiceValue), "--no-requeue")
246 if disp.cluster == nil {
247 // no instance types configured
248 args = append(args, disp.slurmConstraintArgs(container)...)
249 } else if it, err := dispatchcloud.ChooseInstanceType(disp.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured {
251 args = append(args, disp.slurmConstraintArgs(container)...)
252 } else if err != nil {
255 // use instancetype constraint instead of slurm mem/cpu/tmp specs
256 args = append(args, "--constraint=instancetype="+it.Name)
259 if len(container.SchedulingParameters.Partitions) > 0 {
260 args = append(args, "--partition="+strings.Join(container.SchedulingParameters.Partitions, ","))
266 func (disp *Dispatcher) submit(container arvados.Container, crunchRunCommand []string) error {
267 // append() here avoids modifying crunchRunCommand's
268 // underlying array, which is shared with other goroutines.
269 crArgs := append([]string(nil), crunchRunCommand...)
270 crArgs = append(crArgs, container.UUID)
271 crScript := strings.NewReader(execScript(crArgs))
273 sbArgs, err := disp.sbatchArgs(container)
277 log.Printf("running sbatch %+q", sbArgs)
278 return disp.slurm.Batch(crScript, sbArgs)
281 // Submit a container to the slurm queue (or resume monitoring if it's
282 // already in the queue). Cancel the slurm job if the container's
283 // priority changes to zero or its state indicates it's no longer
285 func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
286 ctx, cancel := context.WithCancel(context.Background())
289 if ctr.State == dispatch.Locked && !disp.sqCheck.HasUUID(ctr.UUID) {
290 log.Printf("Submitting container %s to slurm", ctr.UUID)
291 if err := disp.submit(ctr, disp.CrunchRunCommand); err != nil {
293 if err, ok := err.(dispatchcloud.ConstraintsNotSatisfiableError); ok {
294 var logBuf bytes.Buffer
295 fmt.Fprintf(&logBuf, "cannot run container %s: %s\n", ctr.UUID, err)
296 if len(err.AvailableTypes) == 0 {
297 fmt.Fprint(&logBuf, "No instance types are configured.\n")
299 fmt.Fprint(&logBuf, "Available instance types:\n")
300 for _, t := range err.AvailableTypes {
302 "Type %q: %d VCPUs, %d RAM, %d Scratch, %f Price\n",
303 t.Name, t.VCPUs, t.RAM, t.Scratch, t.Price,
307 text = logBuf.String()
308 disp.UpdateState(ctr.UUID, dispatch.Cancelled)
310 text = fmt.Sprintf("Error submitting container %s to slurm: %s", ctr.UUID, err)
314 lr := arvadosclient.Dict{"log": arvadosclient.Dict{
315 "object_uuid": ctr.UUID,
316 "event_type": "dispatch",
317 "properties": map[string]string{"text": text}}}
318 disp.Arv.Create("logs", lr, nil)
320 disp.Unlock(ctr.UUID)
325 log.Printf("Start monitoring container %v in state %q", ctr.UUID, ctr.State)
326 defer log.Printf("Done monitoring container %s", ctr.UUID)
328 // If the container disappears from the slurm queue, there is
329 // no point in waiting for further dispatch updates: just
330 // clean up and return.
331 go func(uuid string) {
332 for ctx.Err() == nil && disp.sqCheck.HasUUID(uuid) {
340 // Disappeared from squeue
341 if err := disp.Arv.Get("containers", ctr.UUID, nil, &ctr); err != nil {
342 log.Printf("error getting final container state for %s: %s", ctr.UUID, err)
345 case dispatch.Running:
346 disp.UpdateState(ctr.UUID, dispatch.Cancelled)
347 case dispatch.Locked:
348 disp.Unlock(ctr.UUID)
351 case updated, ok := <-status:
353 log.Printf("container %s is done: cancel slurm job", ctr.UUID)
355 } else if updated.Priority == 0 {
356 log.Printf("container %s has state %q, priority %d: cancel slurm job", ctr.UUID, updated.State, updated.Priority)
359 p := int64(updated.Priority)
362 // user-assigned priority. If
363 // ctrs have equal priority,
364 // run the older one first.
365 p = int64(p)<<50 - (updated.CreatedAt.UnixNano() >> 14)
367 disp.sqCheck.SetPriority(ctr.UUID, p)
372 func (disp *Dispatcher) scancel(ctr arvados.Container) {
373 err := disp.slurm.Cancel(ctr.UUID)
375 log.Printf("scancel: %s", err)
376 time.Sleep(time.Second)
377 } else if disp.sqCheck.HasUUID(ctr.UUID) {
378 log.Printf("container %s is still in squeue after scancel", ctr.UUID)
379 time.Sleep(time.Second)
383 func (disp *Dispatcher) readConfig(path string) error {
384 err := config.LoadFile(disp, path)
385 if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
386 log.Printf("Config not specified. Continue with default configuration.")