X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/a0f18645f8eccc1f260dfdc71f40ee30a77f75b3..443a0b96316ed46600dc5035193adae6ac4d1f74:/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 9e3baab950..31329c1239 100644 --- a/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go +++ b/services/crunch-dispatch-slurm/crunch-dispatch-slurm.go @@ -7,6 +7,7 @@ package main // Dispatcher service for Crunch that submits containers to the slurm queue. import ( + "bytes" "context" "flag" "fmt" @@ -22,9 +23,15 @@ 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" + "github.com/sirupsen/logrus" "github.com/coreos/go-systemd/daemon" ) +type logger interface { + dispatch.Logger + Fatalf(string, ...interface{}) +} + const initialNiceValue int64 = 10000 var ( @@ -34,6 +41,7 @@ var ( type Dispatcher struct { *dispatch.Dispatcher + logger logrus.FieldLogger cluster *arvados.Cluster sqCheck *SqueueChecker slurm Slurm @@ -56,13 +64,23 @@ type Dispatcher struct { // Minimum time between two attempts to run the same container MinRetryPeriod arvados.Duration + + // Batch size for container queries + BatchSize int64 } func main() { - disp := &Dispatcher{} + logger := logrus.StandardLogger() + if os.Getenv("DEBUG") != "" { + logger.SetLevel(logrus.DebugLevel) + } + logger.Formatter = &logrus.JSONFormatter{ + TimestampFormat: "2006-01-02T15:04:05.000000000Z07:00", + } + disp := &Dispatcher{logger: logger} err := disp.Run(os.Args[0], os.Args[1:]) if err != nil { - log.Fatal(err) + logrus.Fatalf("%s", err) } } @@ -100,7 +118,7 @@ func (disp *Dispatcher) configure(prog string, args []string) error { return nil } - log.Printf("crunch-dispatch-slurm %s started", version) + disp.logger.Printf("crunch-dispatch-slurm %s started", version) err := disp.readConfig(*configPath) if err != nil { @@ -128,7 +146,7 @@ func (disp *Dispatcher) configure(prog string, args []string) error { 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).") + disp.logger.Warnf("Client credentials missing from config, so falling back on environment variables (deprecated).") } if *dumpConfig { @@ -137,7 +155,7 @@ func (disp *Dispatcher) configure(prog string, args []string) error { siteConfig, err := arvados.GetConfig(arvados.DefaultConfigFile) if os.IsNotExist(err) { - log.Printf("warning: no cluster config (%s), proceeding with no node types defined", err) + disp.logger.Warnf("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 { @@ -149,20 +167,26 @@ func (disp *Dispatcher) configure(prog string, args []string) error { // setup() initializes private fields after configure(). func (disp *Dispatcher) setup() { + if disp.logger == nil { + disp.logger = logrus.StandardLogger() + } arv, err := arvadosclient.MakeArvadosClient() if err != nil { - log.Fatalf("Error making Arvados client: %v", err) + disp.logger.Fatalf("Error making Arvados client: %v", err) } arv.Retries = 25 - disp.slurm = &slurmCLI{} + disp.slurm = NewSlurmCLI() disp.sqCheck = &SqueueChecker{ + Logger: disp.logger, Period: time.Duration(disp.PollPeriod), PrioritySpread: disp.PrioritySpread, Slurm: disp.slurm, } disp.Dispatcher = &dispatch.Dispatcher{ Arv: arv, + Logger: disp.logger, + BatchSize: disp.BatchSize, RunContainer: disp.runContainer, PollPeriod: time.Duration(disp.PollPeriod), MinRetryPeriod: time.Duration(disp.MinRetryPeriod), @@ -173,7 +197,7 @@ func (disp *Dispatcher) run() error { defer disp.sqCheck.Stop() if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 { - go dispatchcloud.SlurmNodeTypeFeatureKludge(disp.cluster) + go SlurmNodeTypeFeatureKludge(disp.cluster) } if _, err := daemon.SdNotify(false, "READY=1"); err != nil { @@ -205,12 +229,7 @@ func (disp *Dispatcher) checkSqueueForOrphans() { func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []string { mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM+disp.ReserveExtraRAM) / float64(1048576))) - var disk int64 - for _, m := range container.Mounts { - if m.Kind == "tmp" { - disk += m.Capacity - } - } + disk := dispatchcloud.EstimateScratchSpace(&container) disk = int64(math.Ceil(float64(disk) / float64(1048576))) return []string{ fmt.Sprintf("--mem=%d", mem), @@ -222,7 +241,7 @@ func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []strin 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)) + args = append(args, "--job-name="+container.UUID, fmt.Sprintf("--nice=%d", initialNiceValue), "--no-requeue") if disp.cluster == nil { // no instance types configured @@ -251,9 +270,6 @@ func (disp *Dispatcher) submit(container arvados.Container, crunchRunCommand []s crArgs = append(crArgs, container.UUID) crScript := strings.NewReader(execScript(crArgs)) - disp.sqCheck.L.Lock() - defer disp.sqCheck.L.Unlock() - sbArgs, err := disp.sbatchArgs(container) if err != nil { return err @@ -274,8 +290,21 @@ func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Contain log.Printf("Submitting container %s to slurm", ctr.UUID) if err := disp.submit(ctr, disp.CrunchRunCommand); err != nil { var text string - if err == dispatchcloud.ErrConstraintsNotSatisfiable { - text = fmt.Sprintf("cannot run container %s: %s", ctr.UUID, err) + 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) @@ -310,7 +339,7 @@ func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Contain 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) + log.Printf("error getting final container state for %s: %s", ctr.UUID, err) } switch ctr.State { case dispatch.Running: @@ -341,10 +370,7 @@ func (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Contain } } func (disp *Dispatcher) scancel(ctr arvados.Container) { - disp.sqCheck.L.Lock() err := disp.slurm.Cancel(ctr.UUID) - disp.sqCheck.L.Unlock() - if err != nil { log.Printf("scancel: %s", err) time.Sleep(time.Second)