X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/0eb72b526bf8bbb011551ecf019f604e17a534f1..10f08c358c12468119dc2621c48b68d6d33417da:/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go diff --git a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go index 30cbb79dc1..879cb785d9 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -7,14 +7,12 @@ package main // Dispatcher service for Crunch that submits containers to the slurm queue. import ( - "bytes" "context" "flag" "fmt" "log" "math" "os" - "os/exec" "regexp" "strings" "time" @@ -23,11 +21,18 @@ import ( "git.curoverse.com/arvados.git/sdk/go/arvadosclient" "git.curoverse.com/arvados.git/sdk/go/config" "git.curoverse.com/arvados.git/sdk/go/dispatch" + "git.curoverse.com/arvados.git/services/dispatchcloud" "github.com/coreos/go-systemd/daemon" ) -// Config used by crunch-dispatch-slurm -type Config struct { +var version = "dev" + +type command struct { + dispatcher *dispatch.Dispatcher + cluster *arvados.Cluster + sqCheck *SqueueChecker + slurm Slurm + Client arvados.Client SbatchArguments []string @@ -44,21 +49,16 @@ type Config struct { } func main() { - err := doMain() + err := (&command{}).Run(os.Args[0], os.Args[1:]) if err != nil { log.Fatal(err) } } -var ( - theConfig Config - sqCheck = &SqueueChecker{} -) - const defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml" -func doMain() error { - flags := flag.NewFlagSet("crunch-dispatch-slurm", flag.ExitOnError) +func (cmd *command) Run(prog string, args []string) error { + flags := flag.NewFlagSet(prog, flag.ExitOnError) flags.Usage = func() { usage(flags) } configPath := flags.String( @@ -69,41 +69,52 @@ func doMain() error { "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(os.Args[1:]) + flags.Parse(args) + + // Print version information if requested + if *getVersion { + fmt.Printf("crunch-dispatch-slurm %s\n", version) + return nil + } + + log.Printf("crunch-dispatch-slurm %s started", version) - err := readConfig(&theConfig, *configPath) + err := cmd.readConfig(*configPath) if err != nil { return err } - if theConfig.CrunchRunCommand == nil { - theConfig.CrunchRunCommand = []string{"crunch-run"} + if cmd.CrunchRunCommand == nil { + cmd.CrunchRunCommand = []string{"crunch-run"} } - if theConfig.PollPeriod == 0 { - theConfig.PollPeriod = arvados.Duration(10 * time.Second) + if cmd.PollPeriod == 0 { + cmd.PollPeriod = arvados.Duration(10 * time.Second) } - if theConfig.Client.APIHost != "" || theConfig.Client.AuthToken != "" { + if cmd.Client.APIHost != "" || cmd.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", theConfig.Client.APIHost) - os.Setenv("ARVADOS_API_TOKEN", theConfig.Client.AuthToken) + os.Setenv("ARVADOS_API_HOST", cmd.Client.APIHost) + os.Setenv("ARVADOS_API_TOKEN", cmd.Client.AuthToken) os.Setenv("ARVADOS_API_HOST_INSECURE", "") - if theConfig.Client.Insecure { + if cmd.Client.Insecure { os.Setenv("ARVADOS_API_HOST_INSECURE", "1") } - os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(theConfig.Client.KeepServiceURIs, " ")) + os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(cmd.Client.KeepServiceURIs, " ")) os.Setenv("ARVADOS_EXTERNAL_CLIENT", "") } else { log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).") } if *dumpConfig { - log.Fatal(config.DumpAndExit(theConfig)) + log.Fatal(config.DumpAndExit(cmd)) } arv, err := arvadosclient.MakeArvadosClient() @@ -113,23 +124,41 @@ func doMain() error { } arv.Retries = 25 - sqCheck = &SqueueChecker{Period: time.Duration(theConfig.PollPeriod)} - defer sqCheck.Stop() + siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile) + if os.IsNotExist(err) { + log.Printf("warning: no cluster config file %q (%s), proceeding with no node types defined", arvados.DefaultConfigFile, err) + } else if err != nil { + log.Fatalf("error loading config: %s", err) + } else if cmd.cluster, err = siteConfig.GetCluster(""); err != nil { + log.Fatalf("config error: %s", err) + } else if len(cmd.cluster.InstanceTypes) > 0 { + go dispatchcloud.SlurmNodeTypeFeatureKludge(cmd.cluster) + } + + if cmd.slurm == nil { + cmd.slurm = &slurmCLI{} + } + + cmd.sqCheck = &SqueueChecker{ + Period: time.Duration(cmd.PollPeriod), + Slurm: cmd.slurm, + } + defer cmd.sqCheck.Stop() - dispatcher := &dispatch.Dispatcher{ + cmd.dispatcher = &dispatch.Dispatcher{ Arv: arv, - RunContainer: run, - PollPeriod: time.Duration(theConfig.PollPeriod), - MinRetryPeriod: time.Duration(theConfig.MinRetryPeriod), + RunContainer: cmd.run, + PollPeriod: time.Duration(cmd.PollPeriod), + MinRetryPeriod: time.Duration(cmd.MinRetryPeriod), } if _, err := daemon.SdNotify(false, "READY=1"); err != nil { log.Printf("Error notifying init daemon: %v", err) } - go checkSqueueForOrphans(dispatcher, sqCheck) + go cmd.checkSqueueForOrphans() - return dispatcher.Run(context.Background()) + return cmd.dispatcher.Run(context.Background()) } var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$`) @@ -139,20 +168,30 @@ var containerUuidPattern = regexp.MustCompile(`^[a-z0-9]{5}-dz642-[a-z0-9]{15}$` // 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 checkSqueueForOrphans(dispatcher *dispatch.Dispatcher, sqCheck *SqueueChecker) { - for _, uuid := range sqCheck.All() { +func (cmd *command) checkSqueueForOrphans() { + for _, uuid := range cmd.sqCheck.All() { if !containerUuidPattern.MatchString(uuid) { continue } - err := dispatcher.TrackContainer(uuid) + err := cmd.dispatcher.TrackContainer(uuid) if err != nil { log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err) } } } -// sbatchCmd -func sbatchFunc(container arvados.Container) *exec.Cmd { +func (cmd *command) niceness(priority int) int { + if priority > 1000 { + priority = 1000 + } + if priority < 0 { + priority = 0 + } + // Niceness range 1-10000 + return (1000 - priority) * 10 +} + +func (cmd *command) sbatchArgs(container arvados.Container) ([]string, error) { mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM) / float64(1048576))) var disk int64 @@ -164,82 +203,74 @@ func sbatchFunc(container arvados.Container) *exec.Cmd { disk = int64(math.Ceil(float64(disk) / float64(1048576))) var sbatchArgs []string - sbatchArgs = append(sbatchArgs, theConfig.SbatchArguments...) + sbatchArgs = append(sbatchArgs, cmd.SbatchArguments...) sbatchArgs = append(sbatchArgs, fmt.Sprintf("--job-name=%s", container.UUID)) sbatchArgs = append(sbatchArgs, fmt.Sprintf("--mem=%d", mem)) sbatchArgs = append(sbatchArgs, fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs)) sbatchArgs = append(sbatchArgs, fmt.Sprintf("--tmp=%d", disk)) + sbatchArgs = append(sbatchArgs, fmt.Sprintf("--nice=%d", cmd.niceness(container.Priority))) if len(container.SchedulingParameters.Partitions) > 0 { sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ","))) } - return exec.Command("sbatch", sbatchArgs...) -} + if cmd.cluster == nil { + // no instance types configured + } else if it, err := dispatchcloud.ChooseInstanceType(cmd.cluster, &container); err == dispatchcloud.ErrInstanceTypesNotConfigured { + // ditto + } else if err != nil { + return nil, err + } else { + sbatchArgs = append(sbatchArgs, "--constraint="+it.Name) + } -// scancelCmd -func scancelFunc(container arvados.Container) *exec.Cmd { - return exec.Command("scancel", "--name="+container.UUID) + return sbatchArgs, nil } -// Wrap these so that they can be overridden by tests -var sbatchCmd = sbatchFunc -var scancelCmd = scancelFunc - -// Submit job to slurm using sbatch. -func submit(dispatcher *dispatch.Dispatcher, container arvados.Container, crunchRunCommand []string) error { - cmd := sbatchCmd(container) - - // Send a tiny script on stdin to execute the crunch-run - // command (slurm requires this to be a #! script) - cmd.Stdin = strings.NewReader(execScript(append(crunchRunCommand, container.UUID))) - - var stdout, stderr bytes.Buffer - cmd.Stdout = &stdout - cmd.Stderr = &stderr +func (cmd *command) 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)) - // Mutex between squeue sync and running sbatch or scancel. - sqCheck.L.Lock() - defer sqCheck.L.Unlock() + cmd.sqCheck.L.Lock() + defer cmd.sqCheck.L.Unlock() - log.Printf("exec sbatch %+q", cmd.Args) - err := cmd.Run() - - switch err.(type) { - case nil: - log.Printf("sbatch succeeded: %q", strings.TrimSpace(stdout.String())) - return nil - - case *exec.ExitError: - dispatcher.Unlock(container.UUID) - return fmt.Errorf("sbatch %+q failed: %v (stderr: %q)", cmd.Args, err, stderr.Bytes()) - - default: - dispatcher.Unlock(container.UUID) - return fmt.Errorf("exec failed: %v", err) + sbArgs, err := cmd.sbatchArgs(container) + if err != nil { + return err } + log.Printf("running sbatch %+q", sbArgs) + return cmd.slurm.Batch(crScript, sbArgs) } // 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 run(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) { +func (cmd *command) run(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) { ctx, cancel := context.WithCancel(context.Background()) defer cancel() - if ctr.State == dispatch.Locked && !sqCheck.HasUUID(ctr.UUID) { + if ctr.State == dispatch.Locked && !cmd.sqCheck.HasUUID(ctr.UUID) { log.Printf("Submitting container %s to slurm", ctr.UUID) - if err := submit(disp, ctr, theConfig.CrunchRunCommand); err != nil { - text := fmt.Sprintf("Error submitting container %s to slurm: %s", ctr.UUID, err) + if err := cmd.submit(ctr, cmd.CrunchRunCommand); err != nil { + var text string + if err == dispatchcloud.ErrConstraintsNotSatisfiable { + text = fmt.Sprintf("cannot run container %s: %s", ctr.UUID, err) + cmd.dispatcher.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) + cmd.dispatcher.Arv.Create("logs", lr, nil) - disp.Unlock(ctr.UUID) + cmd.dispatcher.Unlock(ctr.UUID) return } } @@ -251,7 +282,7 @@ func run(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados // no point in waiting for further dispatch updates: just // clean up and return. go func(uuid string) { - for ctx.Err() == nil && sqCheck.HasUUID(uuid) { + for ctx.Err() == nil && cmd.sqCheck.HasUUID(uuid) { } cancel() }(ctr.UUID) @@ -260,45 +291,68 @@ func run(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados select { case <-ctx.Done(): // Disappeared from squeue - if err := disp.Arv.Get("containers", ctr.UUID, nil, &ctr); err != nil { + if err := cmd.dispatcher.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) + cmd.dispatcher.UpdateState(ctr.UUID, dispatch.Cancelled) case dispatch.Locked: - disp.Unlock(ctr.UUID) + cmd.dispatcher.Unlock(ctr.UUID) } return case updated, ok := <-status: if !ok { log.Printf("Dispatcher says container %s is done: cancel slurm job", ctr.UUID) - scancel(ctr) + cmd.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) - scancel(ctr) + cmd.scancel(ctr) + } else { + cmd.renice(updated) } } } } -func scancel(ctr arvados.Container) { - sqCheck.L.Lock() - cmd := scancelCmd(ctr) - msg, err := cmd.CombinedOutput() - sqCheck.L.Unlock() +func (cmd *command) scancel(ctr arvados.Container) { + cmd.sqCheck.L.Lock() + err := cmd.slurm.Cancel(ctr.UUID) + cmd.sqCheck.L.Unlock() if err != nil { - log.Printf("%q %q: %s %q", cmd.Path, cmd.Args, err, msg) + log.Printf("scancel: %s", err) time.Sleep(time.Second) - } else if sqCheck.HasUUID(ctr.UUID) { + } else if cmd.sqCheck.HasUUID(ctr.UUID) { log.Printf("container %s is still in squeue after scancel", ctr.UUID) time.Sleep(time.Second) } } -func readConfig(dst interface{}, path string) error { - err := config.LoadFile(dst, path) +func (cmd *command) renice(ctr arvados.Container) { + nice := cmd.niceness(ctr.Priority) + oldnice := cmd.sqCheck.GetNiceness(ctr.UUID) + if nice == oldnice || oldnice == -1 { + return + } + log.Printf("updating slurm nice value to %d (was %d)", nice, oldnice) + cmd.sqCheck.L.Lock() + err := cmd.slurm.Renice(ctr.UUID, nice) + cmd.sqCheck.L.Unlock() + + if err != nil { + log.Printf("renice: %s", err) + time.Sleep(time.Second) + return + } + if cmd.sqCheck.HasUUID(ctr.UUID) { + log.Printf("container %s has arvados priority %d, slurm nice %d", + ctr.UUID, ctr.Priority, cmd.sqCheck.GetNiceness(ctr.UUID)) + } +} + +func (cmd *command) readConfig(path string) error { + err := config.LoadFile(cmd, path) if err != nil && os.IsNotExist(err) && path == defaultConfigPath { log.Printf("Config not specified. Continue with default configuration.") err = nil