// Dispatcher service for Crunch that submits containers to the slurm queue.
import (
+ "bytes"
"context"
"flag"
"fmt"
"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"
- "git.curoverse.com/arvados.git/services/dispatchcloud"
+ "github.com/sirupsen/logrus"
"github.com/coreos/go-systemd/daemon"
)
+type logger interface {
+ dispatch.Logger
+ Fatalf(string, ...interface{})
+}
+
+const initialNiceValue int64 = 10000
+
var (
version = "dev"
defaultConfigPath = "/etc/arvados/crunch-dispatch-slurm/crunch-dispatch-slurm.yml"
type Dispatcher struct {
*dispatch.Dispatcher
+ logger logrus.FieldLogger
cluster *arvados.Cluster
sqCheck *SqueueChecker
slurm Slurm
SbatchArguments []string
PollPeriod arvados.Duration
+ PrioritySpread int64
// 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
+ // Extra RAM to reserve (in Bytes) for SLURM job, in addition
+ // to the amount specified in the container's RuntimeConstraints
+ ReserveExtraRAM int64
+
// 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)
}
}
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 {
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 {
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 {
// 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{
- Period: time.Duration(disp.PollPeriod),
- Slurm: disp.slurm,
+ 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),
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 {
}
}
-func (disp *Dispatcher) niceness(priority int) int {
- if priority > 1000 {
- priority = 1000
- }
- if priority < 0 {
- priority = 0
+func (disp *Dispatcher) slurmConstraintArgs(container arvados.Container) []string {
+ mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM+disp.ReserveExtraRAM) / float64(1048576)))
+
+ disk := dispatchcloud.EstimateScratchSpace(&container)
+ 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),
}
- // Niceness range 1-10000
- return (1000 - priority) * 10
}
func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) {
- mem := int64(math.Ceil(float64(container.RuntimeConstraints.RAM+container.RuntimeConstraints.KeepCacheRAM) / float64(1048576)))
-
- var disk int64
- for _, m := range container.Mounts {
- if m.Kind == "tmp" {
- disk += m.Capacity
- }
- }
- disk = int64(math.Ceil(float64(disk) / float64(1048576)))
-
- var sbatchArgs []string
- sbatchArgs = append(sbatchArgs, disp.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", disp.niceness(container.Priority)))
- if len(container.SchedulingParameters.Partitions) > 0 {
- sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ",")))
- }
+ var args []string
+ args = append(args, disp.SbatchArguments...)
+ args = append(args, "--job-name="+container.UUID, fmt.Sprintf("--nice=%d", initialNiceValue), "--no-requeue")
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 {
- sbatchArgs = append(sbatchArgs, "--constraint=instancetype="+it.Name)
+ // use instancetype constraint instead of slurm mem/cpu/tmp specs
+ args = append(args, "--constraint=instancetype="+it.Name)
+ }
+
+ if len(container.SchedulingParameters.Partitions) > 0 {
+ args = append(args, "--partition="+strings.Join(container.SchedulingParameters.Partitions, ","))
}
- return sbatchArgs, nil
+ return args, nil
}
func (disp *Dispatcher) submit(container arvados.Container, crunchRunCommand []string) error {
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
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)
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:
log.Printf("container %s has state %q, priority %d: cancel slurm job", ctr.UUID, updated.State, updated.Priority)
disp.scancel(ctr)
} else {
- disp.renice(updated)
+ 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)
}
}
}
}
-
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)
}
}
-func (disp *Dispatcher) renice(ctr arvados.Container) {
- nice := disp.niceness(ctr.Priority)
- oldnice := disp.sqCheck.GetNiceness(ctr.UUID)
- if nice == oldnice || oldnice == -1 {
- return
- }
- log.Printf("updating slurm nice value to %d (was %d)", nice, oldnice)
- disp.sqCheck.L.Lock()
- err := disp.slurm.Renice(ctr.UUID, nice)
- disp.sqCheck.L.Unlock()
-
- if err != nil {
- log.Printf("renice: %s", err)
- time.Sleep(time.Second)
- return
- }
- if disp.sqCheck.HasUUID(ctr.UUID) {
- log.Printf("container %s has arvados priority %d, slurm nice %d",
- ctr.UUID, ctr.Priority, disp.sqCheck.GetNiceness(ctr.UUID))
- }
-}
-
func (disp *Dispatcher) readConfig(path string) error {
err := config.LoadFile(disp, path)
if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
projects / arvados.git / blobdiff