//
// SPDX-License-Identifier: AGPL-3.0
-package main
-
// Dispatcher service for Crunch that submits containers to the slurm queue.
+package dispatchslurm
import (
"context"
- "flag"
+ "crypto/hmac"
+ "crypto/sha256"
"fmt"
"log"
"math"
+ "net/http"
"os"
"regexp"
"strings"
"time"
- "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.arvados.org/arvados.git/lib/cmd"
+ "git.arvados.org/arvados.git/lib/controller/dblock"
+ "git.arvados.org/arvados.git/lib/ctrlctx"
+ "git.arvados.org/arvados.git/lib/dispatchcloud"
+ "git.arvados.org/arvados.git/lib/service"
+ "git.arvados.org/arvados.git/sdk/go/arvados"
+ "git.arvados.org/arvados.git/sdk/go/arvadosclient"
+ "git.arvados.org/arvados.git/sdk/go/ctxlog"
+ "git.arvados.org/arvados.git/sdk/go/dispatch"
"github.com/coreos/go-systemd/daemon"
+ "github.com/prometheus/client_golang/prometheus"
+ "github.com/sirupsen/logrus"
)
-var version = "dev"
+var Command cmd.Handler = service.Command(arvados.ServiceNameDispatchSLURM, newHandler)
-// Config used by crunch-dispatch-slurm
-type Config struct {
- Client arvados.Client
+func newHandler(ctx context.Context, cluster *arvados.Cluster, _ string, _ *prometheus.Registry) service.Handler {
+ logger := ctxlog.FromContext(ctx)
+ disp := &Dispatcher{logger: logger, cluster: cluster}
+ if err := disp.configure(); err != nil {
+ return service.ErrorHandler(ctx, cluster, err)
+ }
+ disp.setup()
+ go func() {
+ disp.err = disp.run()
+ close(disp.done)
+ }()
+ return disp
+}
- SbatchArguments []string
- PollPeriod arvados.Duration
+type logger interface {
+ dispatch.Logger
+ Fatalf(string, ...interface{})
+}
- // 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
+const initialNiceValue int64 = 10000
- // Minimum time between two attempts to run the same container
- MinRetryPeriod arvados.Duration
+type Dispatcher struct {
+ *dispatch.Dispatcher
+ logger logrus.FieldLogger
+ cluster *arvados.Cluster
+ sqCheck *SqueueChecker
+ slurm Slurm
+ dbConnector ctrlctx.DBConnector
- slurm Slurm
-}
+ done chan struct{}
+ err error
-func main() {
- theConfig.slurm = &slurmCLI{}
- err := doMain()
- if err != nil {
- log.Fatal(err)
- }
+ Client arvados.Client
}
-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)
- 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(os.Args[1:])
-
- // Print version information if requested
- if *getVersion {
- fmt.Printf("crunch-dispatch-slurm %s\n", version)
- return nil
- }
+func (disp *Dispatcher) CheckHealth() error {
+ return disp.err
+}
- log.Printf("crunch-dispatch-slurm %s started", version)
+func (disp *Dispatcher) Done() <-chan struct{} {
+ return disp.done
+}
- err := readConfig(&theConfig, *configPath)
- if err != nil {
- return err
- }
+func (disp *Dispatcher) ServeHTTP(w http.ResponseWriter, r *http.Request) {
+ http.NotFound(w, r)
+}
- if theConfig.CrunchRunCommand == nil {
- theConfig.CrunchRunCommand = []string{"crunch-run"}
+// configure() loads config files. Some tests skip this (see
+// StubbedSuite).
+func (disp *Dispatcher) configure() error {
+ if disp.logger == nil {
+ disp.logger = logrus.StandardLogger()
}
+ disp.logger = disp.logger.WithField("ClusterID", disp.cluster.ClusterID)
+ disp.logger.Printf("crunch-dispatch-slurm %s started", cmd.Version.String())
- if theConfig.PollPeriod == 0 {
- theConfig.PollPeriod = arvados.Duration(10 * time.Second)
- }
+ disp.Client.APIHost = disp.cluster.Services.Controller.ExternalURL.Host
+ disp.Client.AuthToken = disp.cluster.SystemRootToken
+ disp.Client.Insecure = disp.cluster.TLS.Insecure
+ disp.dbConnector = ctrlctx.DBConnector{PostgreSQL: disp.cluster.PostgreSQL}
- if theConfig.Client.APIHost != "" || theConfig.Client.AuthToken != "" {
+ 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", theConfig.Client.APIHost)
- os.Setenv("ARVADOS_API_TOKEN", theConfig.Client.AuthToken)
+ os.Setenv("ARVADOS_API_HOST", disp.Client.APIHost)
+ os.Setenv("ARVADOS_API_TOKEN", disp.Client.AuthToken)
os.Setenv("ARVADOS_API_HOST_INSECURE", "")
- if theConfig.Client.Insecure {
+ if disp.Client.Insecure {
os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
}
- os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(theConfig.Client.KeepServiceURIs, " "))
- os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
+ for k, v := range disp.cluster.Containers.SLURM.SbatchEnvironmentVariables {
+ os.Setenv(k, v)
+ }
} else {
- log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).")
- }
-
- if *dumpConfig {
- log.Fatal(config.DumpAndExit(theConfig))
+ disp.logger.Warnf("Client credentials missing from config, so falling back on environment variables (deprecated).")
}
+ return nil
+}
+// setup() initializes private fields after configure().
+func (disp *Dispatcher) setup() {
+ disp.done = make(chan struct{})
arv, err := arvadosclient.MakeArvadosClient()
if err != nil {
- log.Printf("Error making Arvados client: %v", err)
- return err
+ disp.logger.Fatalf("Error making Arvados client: %v", err)
}
arv.Retries = 25
- sqCheck = &SqueueChecker{Period: time.Duration(theConfig.PollPeriod)}
- defer sqCheck.Stop()
-
- dispatcher := &dispatch.Dispatcher{
+ disp.slurm = NewSlurmCLI()
+ disp.sqCheck = &SqueueChecker{
+ Logger: disp.logger,
+ Period: time.Duration(disp.cluster.Containers.CloudVMs.PollInterval),
+ PrioritySpread: disp.cluster.Containers.SLURM.PrioritySpread,
+ Slurm: disp.slurm,
+ }
+ disp.Dispatcher = &dispatch.Dispatcher{
Arv: arv,
- RunContainer: run,
- PollPeriod: time.Duration(theConfig.PollPeriod),
- MinRetryPeriod: time.Duration(theConfig.MinRetryPeriod),
+ Logger: disp.logger,
+ BatchSize: disp.cluster.API.MaxItemsPerResponse,
+ RunContainer: disp.runContainer,
+ PollPeriod: time.Duration(disp.cluster.Containers.CloudVMs.PollInterval),
+ MinRetryPeriod: time.Duration(disp.cluster.Containers.MinRetryPeriod),
+ }
+}
+
+func (disp *Dispatcher) run() error {
+ dblock.Dispatch.Lock(context.Background(), disp.dbConnector.GetDB)
+ defer dblock.Dispatch.Unlock()
+ defer disp.sqCheck.Stop()
+
+ if disp.cluster != nil && len(disp.cluster.InstanceTypes) > 0 {
+ go SlurmNodeTypeFeatureKludge(disp.cluster)
}
if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
log.Printf("Error notifying init daemon: %v", err)
}
-
- go checkSqueueForOrphans(dispatcher, sqCheck)
-
- return dispatcher.Run(context.Background())
+ go disp.checkSqueueForOrphans()
+ return disp.Dispatcher.Run(context.Background())
}
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() {
- if !containerUuidPattern.MatchString(uuid) {
+func (disp *Dispatcher) checkSqueueForOrphans() {
+ for _, uuid := range disp.sqCheck.All() {
+ if !containerUuidPattern.MatchString(uuid) || !strings.HasPrefix(uuid, disp.cluster.ClusterID) {
continue
}
- err := dispatcher.TrackContainer(uuid)
+ err := disp.TrackContainer(uuid)
if err != nil {
log.Printf("checkSqueueForOrphans: TrackContainer(%s): %s", uuid, err)
}
}
}
-func 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+
+ int64(disp.cluster.Containers.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 sbatchArgs(container arvados.Container) []string {
- 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
- }
+func (disp *Dispatcher) sbatchArgs(container arvados.Container) ([]string, error) {
+ var args []string
+ args = append(args, disp.cluster.Containers.SLURM.SbatchArgumentsList...)
+ 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 {
+ // use instancetype constraint instead of slurm mem/cpu/tmp specs
+ args = append(args, "--constraint=instancetype="+it.Name)
}
- disk = int64(math.Ceil(float64(disk) / float64(1048576)))
- var sbatchArgs []string
- sbatchArgs = append(sbatchArgs, theConfig.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", niceness(container.Priority)))
if len(container.SchedulingParameters.Partitions) > 0 {
- sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ",")))
+ args = append(args, "--partition="+strings.Join(container.SchedulingParameters.Partitions, ","))
}
- return sbatchArgs
+ return args, nil
}
-func submit(dispatcher *dispatch.Dispatcher, container arvados.Container, crunchRunCommand []string) error {
+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, "--runtime-engine="+disp.cluster.Containers.RuntimeEngine)
crArgs = append(crArgs, container.UUID)
- crScript := strings.NewReader(execScript(crArgs))
- sqCheck.L.Lock()
- defer sqCheck.L.Unlock()
+ h := hmac.New(sha256.New, []byte(disp.cluster.SystemRootToken))
+ fmt.Fprint(h, container.UUID)
+ authsecret := fmt.Sprintf("%x", h.Sum(nil))
+
+ crScript := strings.NewReader(execScript(crArgs, map[string]string{"GatewayAuthSecret": authsecret}))
- sbArgs := sbatchArgs(container)
+ sbArgs, err := disp.sbatchArgs(container)
+ if err != nil {
+ return err
+ }
log.Printf("running sbatch %+q", sbArgs)
- return theConfig.slurm.Batch(crScript, sbArgs)
+ return disp.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 (disp *Dispatcher) runContainer(_ *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) error {
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
- if ctr.State == dispatch.Locked && !sqCheck.HasUUID(ctr.UUID) {
+ if ctr.State == dispatch.Locked && !disp.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)
- 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
+ cmd := []string{disp.cluster.Containers.CrunchRunCommand}
+ cmd = append(cmd, disp.cluster.Containers.CrunchRunArgumentsList...)
+ err := disp.submit(ctr, cmd)
+ if err != nil {
+ return err
}
}
// 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 && disp.sqCheck.HasUUID(uuid) {
}
cancel()
}(ctr.UUID)
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:
case dispatch.Locked:
disp.Unlock(ctr.UUID)
}
- return
+ return nil
case updated, ok := <-status:
if !ok {
- log.Printf("Dispatcher says container %s is done: cancel slurm job", ctr.UUID)
- scancel(ctr)
+ 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)
- scancel(ctr)
+ log.Printf("container %s has state %q, priority %d: cancel slurm job", ctr.UUID, updated.State, updated.Priority)
+ disp.scancel(ctr)
} else {
- 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 scancel(ctr arvados.Container) {
- sqCheck.L.Lock()
- err := theConfig.slurm.Cancel(ctr.UUID)
- sqCheck.L.Unlock()
-
+func (disp *Dispatcher) scancel(ctr arvados.Container) {
+ err := disp.slurm.Cancel(ctr.UUID)
if err != nil {
log.Printf("scancel: %s", err)
time.Sleep(time.Second)
- } else if sqCheck.HasUUID(ctr.UUID) {
+ } else if disp.sqCheck.HasUUID(ctr.UUID) {
log.Printf("container %s is still in squeue after scancel", ctr.UUID)
time.Sleep(time.Second)
}
}
-
-func renice(ctr arvados.Container) {
- nice := niceness(ctr.Priority)
- oldnice := sqCheck.GetNiceness(ctr.UUID)
- if nice == oldnice || oldnice == -1 {
- return
- }
- log.Printf("updating slurm nice value to %d (was %d)", nice, oldnice)
- sqCheck.L.Lock()
- err := theConfig.slurm.Renice(ctr.UUID, nice)
- sqCheck.L.Unlock()
-
- if err != nil {
- log.Printf("renice: %s", err)
- time.Sleep(time.Second)
- return
- }
- if sqCheck.HasUUID(ctr.UUID) {
- log.Printf("container %s has arvados priority %d, slurm nice %d",
- ctr.UUID, ctr.Priority, sqCheck.GetNiceness(ctr.UUID))
- }
-}
-
-func readConfig(dst interface{}, path string) error {
- err := config.LoadFile(dst, path)
- if err != nil && os.IsNotExist(err) && path == defaultConfigPath {
- log.Printf("Config not specified. Continue with default configuration.")
- err = nil
- }
- return err
-}