3 // Dispatcher service for Crunch that submits containers to the slurm queue.
8 "git.curoverse.com/arvados.git/sdk/go/arvados"
9 "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
10 "git.curoverse.com/arvados.git/sdk/go/config"
11 "git.curoverse.com/arvados.git/sdk/go/dispatch"
12 "github.com/coreos/go-systemd/daemon"
23 // Config used by crunch-dispatch-slurm
27 SbatchArguments []string
28 PollPeriod arvados.Duration
30 // crunch-run command to invoke. The container UUID will be
31 // appended. If nil, []string{"crunch-run"} will be used.
33 // Example: []string{"crunch-run", "--cgroup-parent-subsystem=memory"}
34 CrunchRunCommand []string
49 const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
52 flags := flag.NewFlagSet("crunch-dispatch-slurm", flag.ExitOnError)
53 flags.Usage = func() { usage(flags) }
55 configPath := flags.String(
58 "`path` to JSON or YAML configuration file")
60 // Parse args; omit the first arg which is the command name
61 flags.Parse(os.Args[1:])
63 err := readConfig(&theConfig, *configPath)
68 if theConfig.CrunchRunCommand == nil {
69 theConfig.CrunchRunCommand = []string{"crunch-run"}
72 if theConfig.PollPeriod == 0 {
73 theConfig.PollPeriod = arvados.Duration(10 * time.Second)
76 if theConfig.Client.APIHost != "" || theConfig.Client.AuthToken != "" {
77 // Copy real configs into env vars so [a]
78 // MakeArvadosClient() uses them, and [b] they get
79 // propagated to crunch-run via SLURM.
80 os.Setenv("ARVADOS_API_HOST", theConfig.Client.APIHost)
81 os.Setenv("ARVADOS_API_TOKEN", theConfig.Client.AuthToken)
82 os.Setenv("ARVADOS_API_HOST_INSECURE", "")
83 if theConfig.Client.Insecure {
84 os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
86 os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(theConfig.Client.KeepServiceURIs, " "))
87 os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
89 log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).")
92 arv, err := arvadosclient.MakeArvadosClient()
94 log.Printf("Error making Arvados client: %v", err)
99 squeueUpdater.StartMonitor(time.Duration(theConfig.PollPeriod))
100 defer squeueUpdater.Done()
102 dispatcher := dispatch.Dispatcher{
105 PollInterval: time.Duration(theConfig.PollPeriod),
106 DoneProcessing: make(chan struct{})}
108 if _, err := daemon.SdNotify("READY=1"); err != nil {
109 log.Printf("Error notifying init daemon: %v", err)
112 err = dispatcher.RunDispatcher()
121 func sbatchFunc(container arvados.Container) *exec.Cmd {
122 memPerCPU := math.Ceil(float64(container.RuntimeConstraints.RAM) / (float64(container.RuntimeConstraints.VCPUs) * 1048576))
124 var sbatchArgs []string
125 sbatchArgs = append(sbatchArgs, "--share")
126 sbatchArgs = append(sbatchArgs, theConfig.SbatchArguments...)
127 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--job-name=%s", container.UUID))
128 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--mem-per-cpu=%d", int(memPerCPU)))
129 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs))
130 if container.RuntimeConstraints.Partition != nil {
131 sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.RuntimeConstraints.Partition, ",")))
134 return exec.Command("sbatch", sbatchArgs...)
138 func scancelFunc(container arvados.Container) *exec.Cmd {
139 return exec.Command("scancel", "--name="+container.UUID)
142 // Wrap these so that they can be overridden by tests
143 var sbatchCmd = sbatchFunc
144 var scancelCmd = scancelFunc
146 // Submit job to slurm using sbatch.
147 func submit(dispatcher *dispatch.Dispatcher,
148 container arvados.Container, crunchRunCommand []string) (submitErr error) {
150 // If we didn't get as far as submitting a slurm job,
151 // unlock the container and return it to the queue.
152 if submitErr == nil {
153 // OK, no cleanup needed
156 err := dispatcher.Unlock(container.UUID)
158 log.Printf("Error unlocking container %s: %v", container.UUID, err)
162 // Create the command and attach to stdin/stdout
163 cmd := sbatchCmd(container)
164 stdinWriter, stdinerr := cmd.StdinPipe()
166 submitErr = fmt.Errorf("Error creating stdin pipe %v: %q", container.UUID, stdinerr)
170 stdoutReader, stdoutErr := cmd.StdoutPipe()
171 if stdoutErr != nil {
172 submitErr = fmt.Errorf("Error creating stdout pipe %v: %q", container.UUID, stdoutErr)
176 stderrReader, stderrErr := cmd.StderrPipe()
177 if stderrErr != nil {
178 submitErr = fmt.Errorf("Error creating stderr pipe %v: %q", container.UUID, stderrErr)
182 // Mutex between squeue sync and running sbatch or scancel.
183 squeueUpdater.SlurmLock.Lock()
184 defer squeueUpdater.SlurmLock.Unlock()
186 log.Printf("sbatch starting: %+q", cmd.Args)
189 submitErr = fmt.Errorf("Error starting sbatch: %v", err)
193 stdoutChan := make(chan []byte)
195 b, _ := ioutil.ReadAll(stdoutReader)
200 stderrChan := make(chan []byte)
202 b, _ := ioutil.ReadAll(stderrReader)
207 // Send a tiny script on stdin to execute the crunch-run command
208 // slurm actually enforces that this must be a #! script
209 io.WriteString(stdinWriter, execScript(append(crunchRunCommand, container.UUID)))
214 stdoutMsg := <-stdoutChan
215 stderrmsg := <-stderrChan
221 submitErr = fmt.Errorf("Container submission failed: %v: %v (stderr: %q)", cmd.Args, err, stderrmsg)
225 log.Printf("sbatch succeeded: %s", strings.TrimSpace(string(stdoutMsg)))
229 // If the container is marked as Locked, check if it is already in the slurm
230 // queue. If not, submit it.
232 // If the container is marked as Running, check if it is in the slurm queue.
233 // If not, mark it as Cancelled.
234 func monitorSubmitOrCancel(dispatcher *dispatch.Dispatcher, container arvados.Container, monitorDone *bool) {
237 if squeueUpdater.CheckSqueue(container.UUID) {
238 // Found in the queue, so continue monitoring
240 } else if container.State == dispatch.Locked && !submitted {
241 // Not in queue but in Locked state and we haven't
242 // submitted it yet, so submit it.
244 log.Printf("About to submit queued container %v", container.UUID)
246 if err := submit(dispatcher, container, theConfig.CrunchRunCommand); err != nil {
247 log.Printf("Error submitting container %s to slurm: %v",
249 // maybe sbatch is broken, put it back to queued
250 dispatcher.Unlock(container.UUID)
254 // Not in queue and we are not going to submit it.
255 // Refresh the container state. If it is
256 // Complete/Cancelled, do nothing, if it is Locked then
257 // release it back to the Queue, if it is Running then
258 // clean up the record.
260 var con arvados.Container
261 err := dispatcher.Arv.Get("containers", container.UUID, nil, &con)
263 log.Printf("Error getting final container state: %v", err)
267 case dispatch.Locked:
268 log.Printf("Container %s in state %v but missing from slurm queue, changing to %v.",
269 container.UUID, con.State, dispatch.Queued)
270 dispatcher.Unlock(container.UUID)
271 case dispatch.Running:
272 st := dispatch.Cancelled
273 log.Printf("Container %s in state %v but missing from slurm queue, changing to %v.",
274 container.UUID, con.State, st)
275 dispatcher.UpdateState(container.UUID, st)
277 // Container state is Queued, Complete or Cancelled so stop monitoring it.
284 // Run or monitor a container.
286 // Monitor status updates. If the priority changes to zero, cancel the
287 // container using scancel.
288 func run(dispatcher *dispatch.Dispatcher,
289 container arvados.Container,
290 status chan arvados.Container) {
292 log.Printf("Monitoring container %v started", container.UUID)
293 defer log.Printf("Monitoring container %v finished", container.UUID)
296 go monitorSubmitOrCancel(dispatcher, container, &monitorDone)
298 for container = range status {
299 if container.State == dispatch.Locked || container.State == dispatch.Running {
300 if container.Priority == 0 {
301 log.Printf("Canceling container %s", container.UUID)
303 // Mutex between squeue sync and running sbatch or scancel.
304 squeueUpdater.SlurmLock.Lock()
305 err := scancelCmd(container).Run()
306 squeueUpdater.SlurmLock.Unlock()
309 log.Printf("Error stopping container %s with scancel: %v",
311 if squeueUpdater.CheckSqueue(container.UUID) {
312 log.Printf("Container %s is still in squeue after scancel.",
318 err = dispatcher.UpdateState(container.UUID, dispatch.Cancelled)
325 func readConfig(dst interface{}, path string) error {
326 err := config.LoadFile(dst, path)
327 if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
328 log.Printf("Config not specified. Continue with default configuration.")