+// Copyright (C) The Arvados Authors. All rights reserved.
+//
+// SPDX-License-Identifier: AGPL-3.0
+
package main
+// Dispatcher service for Crunch that submits containers to the slurm queue.
+
import (
- "bufio"
+ "bytes"
+ "context"
"flag"
"fmt"
- "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
- "io/ioutil"
"log"
"math"
"os"
- "os/exec"
- "os/signal"
- "strconv"
- "sync"
- "syscall"
+ "regexp"
+ "strings"
"time"
+
+ "git.curoverse.com/arvados.git/lib/dispatchcloud"
+ "git.curoverse.com/arvados.git/sdk/go/arvados"
+ "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
+ "git.curoverse.com/arvados.git/sdk/go/config"
+ "git.curoverse.com/arvados.git/sdk/go/dispatch"
+ "github.com/coreos/go-systemd/daemon"
)
-func main() {
- err := doMain()
- if err != nil {
- log.Fatalf("%q", err)
- }
-}
+const initialNiceValue int64 = 10000
var (
- arv arvadosclient.ArvadosClient
- runningCmds map[string]*exec.Cmd
- runningCmdsMutex sync.Mutex
- waitGroup sync.WaitGroup
- doneProcessing chan bool
- sigChan chan os.Signal
+ version = "dev"
+ defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
)
-func doMain() error {
- flags := flag.NewFlagSet("crunch-dispatch-slurm", flag.ExitOnError)
+type Dispatcher struct {
+ *dispatch.Dispatcher
+ cluster *arvados.Cluster
+ sqCheck *SqueueChecker
+ slurm Slurm
- pollInterval := flags.Int(
- "poll-interval",
- 10,
- "Interval in seconds to poll for queued containers")
+ Client arvados.Client
- priorityPollInterval := flags.Int(
- "container-priority-poll-interval",
- 60,
- "Interval in seconds to check priority of a dispatched container")
+ SbatchArguments []string
+ PollPeriod arvados.Duration
+ PrioritySpread int64
- crunchRunCommand := flags.String(
- "crunch-run-command",
- "/usr/bin/crunch-run",
- "Crunch command to run container")
+ // crunch-run command to invoke. The container UUID will be
+ // appended. If nil, []string{"crunch-run"} will be used.
+ //
+ // Example: []string{"crunch-run", "--cgroup-parent-subsystem=memory"}
+ CrunchRunCommand []string
- finishCommand := flags.String(
- "finish-command",
- "/usr/bin/crunch-finish-slurm.sh",
- "Command to run from strigger when job is finished")
+ // Extra RAM to reserve (in Bytes) for SLURM job, in addition
+ // to the amount specified in the container's RuntimeConstraints
+ ReserveExtraRAM int64
- // Parse args; omit the first arg which is the command name
- flags.Parse(os.Args[1:])
+ // Minimum time between two attempts to run the same container
+ MinRetryPeriod arvados.Duration
- var err error
- arv, err = arvadosclient.MakeArvadosClient()
+ // Batch size for container queries
+ BatchSize int64
+}
+
+func main() {
+ disp := &Dispatcher{}
+ err := disp.Run(os.Args[0], os.Args[1:])
if err != nil {
- return err
+ log.Fatal(err)
}
+}
- // Channel to terminate
- doneProcessing = make(chan bool)
-
- // Graceful shutdown
- sigChan = make(chan os.Signal, 1)
- signal.Notify(sigChan, syscall.SIGINT, syscall.SIGTERM, syscall.SIGQUIT)
- go func(sig <-chan os.Signal) {
- for sig := range sig {
- log.Printf("Caught signal: %v", sig)
- doneProcessing <- true
- }
- }(sigChan)
+func (disp *Dispatcher) Run(prog string, args []string) error {
+ if err := disp.configure(prog, args); err != nil {
+ return err
+ }
+ disp.setup()
+ return disp.run()
+}
- // Run all queued containers
- runQueuedContainers(*pollInterval, *priorityPollInterval, *crunchRunCommand, *finishCommand)
+// configure() loads config files. Tests skip this.
+func (disp *Dispatcher) configure(prog string, args []string) error {
+ flags := flag.NewFlagSet(prog, flag.ExitOnError)
+ flags.Usage = func() { usage(flags) }
+
+ configPath := flags.String(
+ "config",
+ defaultConfigPath,
+ "`path` to JSON or YAML configuration file")
+ dumpConfig := flag.Bool(
+ "dump-config",
+ false,
+ "write current configuration to stdout and exit")
+ getVersion := flags.Bool(
+ "version",
+ false,
+ "Print version information and exit.")
+ // Parse args; omit the first arg which is the command name
+ flags.Parse(args)
- // Wait for all running crunch jobs to complete / terminate
- waitGroup.Wait()
+ // Print version information if requested
+ if *getVersion {
+ fmt.Printf("crunch-dispatch-slurm %s\n", version)
+ return nil
+ }
- return nil
-}
+ log.Printf("crunch-dispatch-slurm %s started", version)
-type apiClientAuthorization struct {
- UUID string `json:"uuid"`
- APIToken string `json:"api_token"`
-}
+ err := disp.readConfig(*configPath)
+ if err != nil {
+ return err
+ }
-type apiClientAuthorizationList struct {
- Items []apiClientAuthorization `json:"items"`
-}
+ if disp.CrunchRunCommand == nil {
+ disp.CrunchRunCommand = []string{"crunch-run"}
+ }
-// Poll for queued containers using pollInterval.
-// Invoke dispatchSlurm for each ticker cycle, which will run all the queued containers.
-//
-// Any errors encountered are logged but the program would continue to run (not exit).
-// This is because, once one or more crunch jobs are running,
-// we would need to wait for them complete.
-func runQueuedContainers(pollInterval, priorityPollInterval int, crunchRunCommand, finishCommand string) {
- var auth apiClientAuthorization
- err := arv.Call("GET", "api_client_authorizations", "", "current", nil, &auth)
- if err != nil {
- log.Printf("Error getting my token UUID: %v", err)
- return
+ if disp.PollPeriod == 0 {
+ disp.PollPeriod = arvados.Duration(10 * time.Second)
}
- ticker := time.NewTicker(time.Duration(pollInterval) * time.Second)
- for {
- select {
- case <-ticker.C:
- dispatchSlurm(auth, time.Duration(priorityPollInterval)*time.Second, crunchRunCommand, finishCommand)
- case <-doneProcessing:
- ticker.Stop()
- return
+ if disp.Client.APIHost != "" || disp.Client.AuthToken != "" {
+ // Copy real configs into env vars so [a]
+ // MakeArvadosClient() uses them, and [b] they get
+ // propagated to crunch-run via SLURM.
+ os.Setenv("ARVADOS_API_HOST", disp.Client.APIHost)
+ os.Setenv("ARVADOS_API_TOKEN", disp.Client.AuthToken)
+ os.Setenv("ARVADOS_API_HOST_INSECURE", "")
+ if disp.Client.Insecure {
+ os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
}
+ os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(disp.Client.KeepServiceURIs, " "))
+ os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
+ } else {
+ log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).")
}
-}
-
-// Container data
-type Container struct {
- UUID string `json:"uuid"`
- State string `json:"state"`
- Priority int `json:"priority"`
- RuntimeConstraints map[string]int64 `json:"runtime_constraints"`
- LockedByUUID string `json:"locked_by_uuid"`
-}
-// ContainerList is a list of the containers from api
-type ContainerList struct {
- Items []Container `json:"items"`
-}
+ if *dumpConfig {
+ return config.DumpAndExit(disp)
+ }
-// Get the list of queued containers from API server and invoke run
-// for each container.
-func dispatchSlurm(auth apiClientAuthorization, pollInterval time.Duration, crunchRunCommand, finishCommand string) {
- params := arvadosclient.Dict{
- "filters": [][]interface{}{{"state", "in", []string{"Queued", "Locked"}}},
+ siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile)
+ if os.IsNotExist(err) {
+ log.Printf("warning: no cluster config (%s), proceeding with no node types defined", err)
+ } else if err != nil {
+ return fmt.Errorf("error loading config: %s", err)
+ } else if disp.cluster, err = siteConfig.GetCluster(""); err != nil {
+ return fmt.Errorf("config error: %s", err)
}
- var containers ContainerList
- err := arv.List("containers", params, &containers)
+ return nil
+}
+
+// setup() initializes private fields after configure().
+func (disp *Dispatcher) setup() {
+ arv, err := arvadosclient.MakeArvadosClient()
if err != nil {
- log.Printf("Error getting list of queued containers: %q", err)
- return
+ log.Fatalf("Error making Arvados client: %v", err)
}
+ arv.Retries = 25
- for _, container := range containers.Items {
- if container.State == "Locked" {
- if container.LockedByUUID != auth.UUID {
- // Locked by a different dispatcher
- continue
- } else if checkMine(container.UUID) {
- // I already have a goroutine running
- // for this container: it just hasn't
- // gotten past Locked state yet.
- continue
- }
- log.Printf("WARNING: found container %s already locked by my token %s, but I didn't submit it. "+
- "Assuming it was left behind by a previous dispatch process, and waiting for it to finish.",
- container.UUID, auth.UUID)
- setMine(container.UUID, true)
- go func() {
- waitContainer(container, pollInterval)
- setMine(container.UUID, false)
- }()
- }
- go run(container, crunchRunCommand, finishCommand, pollInterval)
+ disp.slurm = &slurmCLI{}
+ disp.sqCheck = &SqueueChecker{
+ Period: time.Duration(disp.PollPeriod),
+ PrioritySpread: disp.PrioritySpread,
+ Slurm: disp.slurm,
+ }
+ disp.Dispatcher = &dispatch.Dispatcher{
+ Arv: arv,
+ BatchSize: disp.BatchSize,
+ RunContainer: disp.runContainer,
+ PollPeriod: time.Duration(disp.PollPeriod),
+ MinRetryPeriod: time.Duration(disp.MinRetryPeriod),
}
}
-// sbatchCmd
-func sbatchFunc(container Container) *exec.Cmd {
- memPerCPU := math.Ceil((float64(container.RuntimeConstraints["ram"])) / (float64(container.RuntimeConstraints["vcpus"] * 1048576)))
- return exec.Command("sbatch", "--share", "--parsable",
- "--job-name="+container.UUID,
- "--mem-per-cpu="+strconv.Itoa(int(memPerCPU)),
- "--cpus-per-task="+strconv.Itoa(int(container.RuntimeConstraints["vcpus"])))
-}
+func (disp *Dispatcher) run() error {
+ defer disp.sqCheck.Stop()
-var sbatchCmd = sbatchFunc
+ if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 {
+ go dispatchcloud.SlurmNodeTypeFeatureKludge(disp.cluster)
+ }
-// striggerCmd
-func striggerFunc(jobid, containerUUID, finishCommand, apiHost, apiToken, apiInsecure string) *exec.Cmd {
- return exec.Command("strigger", "--set", "--jobid="+jobid, "--fini",
- fmt.Sprintf("--program=%s %s %s %s %s", finishCommand, apiHost, apiToken, apiInsecure, containerUUID))
+ if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
+ log.Printf("Error notifying init daemon: %v", err)
+ }
+ go disp.checkSqueueForOrphans()
+ return disp.Dispatcher.Run(context.Background())
}
-var striggerCmd = striggerFunc
-
-// Submit job to slurm using sbatch.
-func submit(container Container, crunchRunCommand string) (jobid string, submitErr error) {
- submitErr = nil
+var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$`)
- defer func() {
- // If we didn't get as far as submitting a slurm job,
- // unlock the container and return it to the queue.
- if submitErr == nil {
- // OK, no cleanup needed
- return
+// Check the next squeue report, and invoke TrackContainer for all the
+// containers in the report. This gives us a chance to cancel slurm
+// jobs started by a previous dispatch process that never released
+// their slurm allocations even though their container states are
+// Cancelled or Complete. See https://dev.arvados.org/issues/10979
+func (disp *Dispatcher) checkSqueueForOrphans() {
+ for _, uuid := range disp.sqCheck.All() {
+ if !containerUuidPattern.MatchString(uuid) {
+ continue
}
- err := arv.Update("containers", container.UUID,
- arvadosclient.Dict{
- "container": arvadosclient.Dict{"state": "Queued"}},
- nil)
+ err := disp.TrackContainer(uuid)
if err != nil {
- log.Printf("Error unlocking container %s: %v", container.UUID, err)
+ log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err)
}
- }()
-
- // Create the command and attach to stdin/stdout
- cmd := sbatchCmd(container)
- stdinWriter, stdinerr := cmd.StdinPipe()
- if stdinerr != nil {
- submitErr = fmt.Errorf("Error creating stdin pipe %v: %q", container.UUID, stdinerr)
- return
}
+}
- stdoutReader, stdoutErr := cmd.StdoutPipe()
- if stdoutErr != nil {
- submitErr = fmt.Errorf("Error creating stdout pipe %v: %q", container.UUID, stdoutErr)
- return
- }
+func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []string {
+ mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM+disp.ReserveExtraRAM) / float64(1048576)))
- stderrReader, stderrErr := cmd.StderrPipe()
- if stderrErr != nil {
- submitErr = fmt.Errorf("Error creating stderr pipe %v: %q", container.UUID, stderrErr)
- return
- }
-
- err := cmd.Start()
- if err != nil {
- submitErr = fmt.Errorf("Error starting %v: %v", cmd.Args, err)
- return
+ var disk int64
+ for _, m := range container.Mounts {
+ if m.Kind == "tmp" {
+ disk += m.Capacity
+ }
}
-
- stdoutChan := make(chan []byte)
- go func() {
- b, _ := ioutil.ReadAll(stdoutReader)
- stdoutReader.Close()
- stdoutChan <- b
- close(stdoutChan)
- }()
-
- stderrChan := make(chan []byte)
- go func() {
- b, _ := ioutil.ReadAll(stderrReader)
- stderrReader.Close()
- stderrChan <- b
- close(stderrChan)
- }()
-
- // Send a tiny script on stdin to execute the crunch-run command
- // slurm actually enforces that this must be a #! script
- fmt.Fprintf(stdinWriter, "#!/bin/sh\nexec '%s' '%s'\n", crunchRunCommand, container.UUID)
- stdinWriter.Close()
-
- err = cmd.Wait()
-
- stdoutMsg := <-stdoutChan
- stderrmsg := <-stderrChan
-
- if err != nil {
- submitErr = fmt.Errorf("Container submission failed %v: %v %v", cmd.Args, err, stderrmsg)
- return
+ disk = int64(math.Ceil(float64(disk) / float64(1048576)))
+ return []string{
+ fmt.Sprintf("--mem=%d", mem),
+ fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs),
+ fmt.Sprintf("--tmp=%d", disk),
}
-
- // If everything worked out, got the jobid on stdout
- jobid = string(stdoutMsg)
-
- return
}
-// finalizeRecordOnFinish uses 'strigger' command to register a script that will run on
-// the slurm controller when the job finishes.
-func finalizeRecordOnFinish(jobid, containerUUID, finishCommand, apiHost, apiToken, apiInsecure string) {
- cmd := striggerCmd(jobid, containerUUID, finishCommand, apiHost, apiToken, apiInsecure)
- cmd.Stdout = os.Stdout
- cmd.Stderr = os.Stderr
- err := cmd.Run()
- if err != nil {
- log.Printf("While setting up strigger: %v", err)
- // BUG: we drop the error here and forget about it. A
- // human has to notice the container is stuck in
- // Running state, and fix it manually.
+func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) {
+ var args []string
+ args = append(args, disp.SbatchArguments...)
+ args = append(args, "--job-name="+container.UUID, fmt.Sprintf("--nice=%d", initialNiceValue))
+
+ if disp.cluster == nil {
+ // no instance types configured
+ args = append(args, disp.slurmConstraintArgs(container)...)
+ } else if it, err := dispatchcloud.ChooseInstanceType(disp.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured {
+ // ditto
+ args = append(args, disp.slurmConstraintArgs(container)...)
+ } else if err != nil {
+ return nil, err
+ } else {
+ // use instancetype constraint instead of slurm mem/cpu/tmp specs
+ args = append(args, "--constraint=instancetype="+it.Name)
}
-}
-// Run a queued container: [1] Set container state to locked. [2]
-// Execute crunch-run as a slurm batch job. [3] waitContainer().
-func run(container Container, crunchRunCommand, finishCommand string, pollInterval time.Duration) {
- setMine(container.UUID, true)
- defer setMine(container.UUID, false)
-
- // Update container status to Locked. This will fail if
- // another dispatcher (token) has already locked it. It will
- // succeed if *this* dispatcher has already locked it.
- err := arv.Update("containers", container.UUID,
- arvadosclient.Dict{
- "container": arvadosclient.Dict{"state": "Locked"}},
- nil)
- if err != nil {
- log.Printf("Error updating container state to 'Locked' for %v: %q", container.UUID, err)
- return
+ if len(container.SchedulingParameters.Partitions) > 0 {
+ args = append(args, "--partition="+strings.Join(container.SchedulingParameters.Partitions, ","))
}
- log.Printf("About to submit queued container %v", container.UUID)
+ return args, nil
+}
- jobid, err := submit(container, crunchRunCommand)
- if err != nil {
- log.Printf("Error submitting container %s to slurm: %v", container.UUID, err)
- return
- }
+func (disp *Dispatcher) submit(container arvados.Container, crunchRunCommand []string) error {
+ // append() here avoids modifying crunchRunCommand's
+ // underlying array, which is shared with other goroutines.
+ crArgs := append([]string(nil), crunchRunCommand...)
+ crArgs = append(crArgs, container.UUID)
+ crScript := strings.NewReader(execScript(crArgs))
- insecure := "0"
- if arv.ApiInsecure {
- insecure = "1"
- }
- finalizeRecordOnFinish(jobid, container.UUID, finishCommand, arv.ApiServer, arv.ApiToken, insecure)
-
- // Update container status to Running. This will fail if
- // another dispatcher (token) has already locked it. It will
- // succeed if *this* dispatcher has already locked it.
- err = arv.Update("containers", container.UUID,
- arvadosclient.Dict{
- "container": arvadosclient.Dict{"state": "Running"}},
- nil)
+ disp.sqCheck.L.Lock()
+ defer disp.sqCheck.L.Unlock()
+
+ sbArgs, err := disp.sbatchArgs(container)
if err != nil {
- log.Printf("Error updating container state to 'Running' for %v: %q", container.UUID, err)
+ return err
}
- log.Printf("Submitted container %v to slurm", container.UUID)
- waitContainer(container, pollInterval)
+ log.Printf("running sbatch %+q", sbArgs)
+ return disp.slurm.Batch(crScript, sbArgs)
}
-// Wait for a container to finish. Cancel the slurm job if the
-// container priority changes to zero before it ends.
-func waitContainer(container Container, pollInterval time.Duration) {
- log.Printf("Monitoring container %v started", container.UUID)
- defer log.Printf("Monitoring container %v finished", container.UUID)
-
- pollTicker := time.NewTicker(pollInterval)
- defer pollTicker.Stop()
- for _ = range pollTicker.C {
- var updated Container
- err := arv.Get("containers", container.UUID, nil, &updated)
- if err != nil {
- log.Printf("Error getting container %s: %q", container.UUID, err)
- continue
- }
- if updated.State == "Complete" || updated.State == "Cancelled" {
+// Submit a container to the slurm queue (or resume monitoring if it's
+// already in the queue). Cancel the slurm job if the container's
+// priority changes to zero or its state indicates it's no longer
+// running.
+func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
+ ctx, cancel := context.WithCancel(context.Background())
+ defer cancel()
+
+ if ctr.State == dispatch.Locked && !disp.sqCheck.HasUUID(ctr.UUID) {
+ log.Printf("Submitting container %s to slurm", ctr.UUID)
+ if err := disp.submit(ctr, disp.CrunchRunCommand); err != nil {
+ var text string
+ if err, ok := err.(dispatchcloud.ConstraintsNotSatisfiableError); ok {
+ var logBuf bytes.Buffer
+ fmt.Fprintf(&logBuf, "cannot run container %s: %s\n", ctr.UUID, err)
+ if len(err.AvailableTypes) == 0 {
+ fmt.Fprint(&logBuf, "No instance types are configured.\n")
+ } else {
+ fmt.Fprint(&logBuf, "Available instance types:\n")
+ for _, t := range err.AvailableTypes {
+ fmt.Fprintf(&logBuf,
+ "Type %q: %d VCPUs, %d RAM, %d Scratch, %f Price\n",
+ t.Name, t.VCPUs, t.RAM, t.Scratch, t.Price,
+ )
+ }
+ }
+ text = logBuf.String()
+ disp.UpdateState(ctr.UUID, dispatch.Cancelled)
+ } else {
+ text = fmt.Sprintf("Error submitting container %s to slurm: %s", ctr.UUID, err)
+ }
+ log.Print(text)
+
+ lr := arvadosclient.Dict{"log": arvadosclient.Dict{
+ "object_uuid": ctr.UUID,
+ "event_type": "dispatch",
+ "properties": map[string]string{"text": text}}}
+ disp.Arv.Create("logs", lr, nil)
+
+ disp.Unlock(ctr.UUID)
return
}
- if updated.Priority != 0 {
- continue
- }
+ }
- // Priority is zero, but state is Running or Locked
- log.Printf("Canceling container %s", container.UUID)
+ log.Printf("Start monitoring container %v in state %q", ctr.UUID, ctr.State)
+ defer log.Printf("Done monitoring container %s", ctr.UUID)
- err = exec.Command("scancel", "--name="+container.UUID).Run()
- if err != nil {
- log.Printf("Error stopping container %s with scancel: %v", container.UUID, err)
- if inQ, err := checkSqueue(container.UUID); err != nil {
- log.Printf("Error running squeue: %v", err)
- continue
- } else if inQ {
- log.Printf("Container %s is still in squeue; will retry", container.UUID)
- continue
- }
+ // If the container disappears from the slurm queue, there is
+ // no point in waiting for further dispatch updates: just
+ // clean up and return.
+ go func(uuid string) {
+ for ctx.Err() == nil && disp.sqCheck.HasUUID(uuid) {
}
+ cancel()
+ }(ctr.UUID)
- err = arv.Update("containers", container.UUID,
- arvadosclient.Dict{
- "container": arvadosclient.Dict{"state": "Cancelled"}},
- nil)
- if err != nil {
- log.Printf("Error updating state for container %s: %s", container.UUID, err)
- continue
+ for {
+ select {
+ case <-ctx.Done():
+ // Disappeared from squeue
+ if err := disp.Arv.Get("containers", ctr.UUID, nil, &ctr); err != nil {
+ log.Printf("Error getting final container state for %s: %s", ctr.UUID, err)
+ }
+ switch ctr.State {
+ case dispatch.Running:
+ disp.UpdateState(ctr.UUID, dispatch.Cancelled)
+ case dispatch.Locked:
+ disp.Unlock(ctr.UUID)
+ }
+ return
+ case updated, ok := <-status:
+ if !ok {
+ log.Printf("container %s is done: cancel slurm job", ctr.UUID)
+ disp.scancel(ctr)
+ } else if updated.Priority == 0 {
+ log.Printf("container %s has state %q, priority %d: cancel slurm job", ctr.UUID, updated.State, updated.Priority)
+ disp.scancel(ctr)
+ } else {
+ p := int64(updated.Priority)
+ if p <= 1000 {
+ // API is providing
+ // user-assigned priority. If
+ // ctrs have equal priority,
+ // run the older one first.
+ p = int64(p)<<50 - (updated.CreatedAt.UnixNano() >> 14)
+ }
+ disp.sqCheck.SetPriority(ctr.UUID, p)
+ }
}
-
- return
}
}
+func (disp *Dispatcher) scancel(ctr arvados.Container) {
+ disp.sqCheck.L.Lock()
+ err := disp.slurm.Cancel(ctr.UUID)
+ disp.sqCheck.L.Unlock()
-func checkSqueue(uuid string) (bool, error) {
- cmd := exec.Command("squeue", "--format=%j")
- sq, err := cmd.StdoutPipe()
if err != nil {
- return false, err
+ log.Printf("scancel: %s", err)
+ time.Sleep(time.Second)
+ } else if disp.sqCheck.HasUUID(ctr.UUID) {
+ log.Printf("container %s is still in squeue after scancel", ctr.UUID)
+ time.Sleep(time.Second)
}
- cmd.Start()
- defer cmd.Wait()
- scanner := bufio.NewScanner(sq)
- found := false
- for scanner.Scan() {
- if scanner.Text() == uuid {
- found = true
- }
- }
- if err := scanner.Err(); err != nil {
- return false, err
- }
- return found, nil
}
-var mineMutex sync.RWMutex
-var mineMap = make(map[string]bool)
-
-// Goroutine-safely add/remove uuid to the set of "my" containers,
-// i.e., ones for which this process has a goroutine running.
-func setMine(uuid string, t bool) {
- mineMutex.Lock()
- if t {
- mineMap[uuid] = true
- } else {
- delete(mineMap, uuid)
+func (disp *Dispatcher) readConfig(path string) error {
+ err := config.LoadFile(disp, path)
+ if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
+ log.Printf("Config not specified. Continue with default configuration.")
+ err = nil
}
- mineMutex.Unlock()
-}
-
-// Check whether there is already a goroutine running for this
-// container.
-func checkMine(uuid string) bool {
- mineMutex.RLocker().Lock()
- defer mineMutex.RLocker().Unlock()
- return mineMap[uuid]
+ return err
}
projects / arvados.git / blobdiff