import (
"flag"
"fmt"
+ "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"
+ "io"
"io/ioutil"
"log"
"math"
"time"
)
+// Config used by crunch-dispatch-slurm
+type Config struct {
+ Client arvados.Client
+
+ SbatchArguments []string
+ PollPeriod arvados.Duration
+
+ // 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
+}
+
func main() {
err := doMain()
if err != nil {
- log.Fatalf("%q", err)
+ log.Fatal(err)
}
}
var (
- crunchRunCommand *string
- squeueUpdater Squeue
+ theConfig Config
+ squeueUpdater Squeue
)
+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) }
- pollInterval := flags.Int(
- "poll-interval",
- 10,
- "Interval in seconds to poll for queued containers")
-
- crunchRunCommand = flags.String(
- "crunch-run-command",
- "/usr/bin/crunch-run",
- "Crunch command to run container")
+ configPath := flags.String(
+ "config",
+ defaultConfigPath,
+ "`path` to JSON or YAML configuration file")
// Parse args; omit the first arg which is the command name
flags.Parse(os.Args[1:])
+ err := readConfig(&theConfig, *configPath)
+ if err != nil {
+ return err
+ }
+
+ if theConfig.CrunchRunCommand == nil {
+ theConfig.CrunchRunCommand = []string{"crunch-run"}
+ }
+
+ if theConfig.PollPeriod == 0 {
+ theConfig.PollPeriod = arvados.Duration(10 * time.Second)
+ }
+
+ if theConfig.Client.APIHost != "" || theConfig.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_INSECURE", "")
+ if theConfig.Client.Insecure {
+ os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
+ }
+ os.Setenv("ARVADOS_KEEP_SERVICES", strings.Join(theConfig.Client.KeepServiceURIs, " "))
+ os.Setenv("ARVADOS_EXTERNAL_CLIENT", "")
+ } else {
+ log.Printf("warning: Client credentials missing from config, so falling back on environment variables (deprecated).")
+ }
+
arv, err := arvadosclient.MakeArvadosClient()
if err != nil {
log.Printf("Error making Arvados client: %v", err)
}
arv.Retries = 25
- squeueUpdater.StartMonitor(time.Duration(*pollInterval) * time.Second)
+ squeueUpdater.StartMonitor(time.Duration(theConfig.PollPeriod))
defer squeueUpdater.Done()
dispatcher := dispatch.Dispatcher{
Arv: arv,
RunContainer: run,
- PollInterval: time.Duration(*pollInterval) * time.Second,
+ PollInterval: time.Duration(theConfig.PollPeriod),
DoneProcessing: make(chan struct{})}
+ if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
+ log.Printf("Error notifying init daemon: %v", err)
+ }
+
err = dispatcher.RunDispatcher()
if err != nil {
return err
}
// sbatchCmd
-func sbatchFunc(container dispatch.Container) *exec.Cmd {
- memPerCPU := math.Ceil((float64(container.RuntimeConstraints["ram"])) / (float64(container.RuntimeConstraints["vcpus"] * 1048576)))
- return exec.Command("sbatch", "--share", "--parsable",
- fmt.Sprintf("--job-name=%s", container.UUID),
- fmt.Sprintf("--mem-per-cpu=%d", int(memPerCPU)),
- fmt.Sprintf("--cpus-per-task=%d", int(container.RuntimeConstraints["vcpus"])),
- fmt.Sprintf("--priority=%d", container.Priority))
+func sbatchFunc(container arvados.Container) *exec.Cmd {
+ memPerCPU := math.Ceil(float64(container.RuntimeConstraints.RAM) / (float64(container.RuntimeConstraints.VCPUs) * 1048576))
+
+ var sbatchArgs []string
+ sbatchArgs = append(sbatchArgs, "--share")
+ sbatchArgs = append(sbatchArgs, theConfig.SbatchArguments...)
+ sbatchArgs = append(sbatchArgs, fmt.Sprintf("--job-name=%s", container.UUID))
+ sbatchArgs = append(sbatchArgs, fmt.Sprintf("--mem-per-cpu=%d", int(memPerCPU)))
+ sbatchArgs = append(sbatchArgs, fmt.Sprintf("--cpus-per-task=%d", container.RuntimeConstraints.VCPUs))
+ if container.SchedulingParameters.Partitions != nil {
+ sbatchArgs = append(sbatchArgs, fmt.Sprintf("--partition=%s", strings.Join(container.SchedulingParameters.Partitions, ",")))
+ }
+
+ return exec.Command("sbatch", sbatchArgs...)
}
// scancelCmd
-func scancelFunc(container dispatch.Container) *exec.Cmd {
+func scancelFunc(container arvados.Container) *exec.Cmd {
return exec.Command("scancel", "--name="+container.UUID)
}
// Submit job to slurm using sbatch.
func submit(dispatcher *dispatch.Dispatcher,
- container dispatch.Container, crunchRunCommand string) (jobid string, submitErr error) {
- submitErr = nil
-
+ container arvados.Container, crunchRunCommand []string) (submitErr error) {
defer func() {
// If we didn't get as far as submitting a slurm job,
// unlock the container and return it to the queue.
// OK, no cleanup needed
return
}
- err := dispatcher.Arv.Update("containers", container.UUID,
- arvadosclient.Dict{
- "container": arvadosclient.Dict{"state": "Queued"}},
- nil)
+ err := dispatcher.Unlock(container.UUID)
if err != nil {
log.Printf("Error unlocking container %s: %v", container.UUID, err)
}
squeueUpdater.SlurmLock.Lock()
defer squeueUpdater.SlurmLock.Unlock()
+ log.Printf("sbatch starting: %+q", cmd.Args)
err := cmd.Start()
if err != nil {
- submitErr = fmt.Errorf("Error starting %v: %v", cmd.Args, err)
+ submitErr = fmt.Errorf("Error starting sbatch: %v", err)
return
}
b, _ := ioutil.ReadAll(stdoutReader)
stdoutReader.Close()
stdoutChan <- b
+ close(stdoutChan)
}()
stderrChan := make(chan []byte)
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)
+ io.WriteString(stdinWriter, execScript(append(crunchRunCommand, container.UUID)))
stdinWriter.Close()
- err = cmd.Wait()
-
stdoutMsg := <-stdoutChan
stderrmsg := <-stderrChan
- close(stdoutChan)
- close(stderrChan)
+ err = cmd.Wait()
if err != nil {
- submitErr = fmt.Errorf("Container submission failed %v: %v %v", cmd.Args, err, stderrmsg)
+ submitErr = fmt.Errorf("Container submission failed: %v: %v (stderr: %q)", cmd.Args, err, stderrmsg)
return
}
- // If everything worked out, got the jobid on stdout
- jobid = strings.TrimSpace(string(stdoutMsg))
-
+ log.Printf("sbatch succeeded: %s", strings.TrimSpace(string(stdoutMsg)))
return
}
//
// If the container is marked as Running, check if it is in the slurm queue.
// If not, mark it as Cancelled.
-func monitorSubmitOrCancel(dispatcher *dispatch.Dispatcher, container dispatch.Container, monitorDone *bool) {
+func monitorSubmitOrCancel(dispatcher *dispatch.Dispatcher, container arvados.Container, monitorDone *bool) {
submitted := false
for !*monitorDone {
if squeueUpdater.CheckSqueue(container.UUID) {
log.Printf("About to submit queued container %v", container.UUID)
- if _, err := submit(dispatcher, container, *crunchRunCommand); err != nil {
+ if err := submit(dispatcher, container, theConfig.CrunchRunCommand); err != nil {
log.Printf("Error submitting container %s to slurm: %v",
container.UUID, err)
// maybe sbatch is broken, put it back to queued
- dispatcher.UpdateState(container.UUID, dispatch.Queued)
+ dispatcher.Unlock(container.UUID)
}
submitted = true
} else {
// release it back to the Queue, if it is Running then
// clean up the record.
- var con dispatch.Container
+ var con arvados.Container
err := dispatcher.Arv.Get("containers", container.UUID, nil, &con)
if err != nil {
log.Printf("Error getting final container state: %v", err)
}
- var st string
switch con.State {
case dispatch.Locked:
- st = dispatch.Queued
+ log.Printf("Container %s in state %v but missing from slurm queue, changing to %v.",
+ container.UUID, con.State, dispatch.Queued)
+ dispatcher.Unlock(container.UUID)
case dispatch.Running:
- st = dispatch.Cancelled
+ st := dispatch.Cancelled
+ log.Printf("Container %s in state %v but missing from slurm queue, changing to %v.",
+ container.UUID, con.State, st)
+ dispatcher.UpdateState(container.UUID, st)
default:
// Container state is Queued, Complete or Cancelled so stop monitoring it.
return
}
-
- log.Printf("Container %s in state %v but missing from slurm queue, changing to %v.",
- container.UUID, con.State, st)
- dispatcher.UpdateState(container.UUID, st)
}
}
}
// Monitor status updates. If the priority changes to zero, cancel the
// container using scancel.
func run(dispatcher *dispatch.Dispatcher,
- container dispatch.Container,
- status chan dispatch.Container) {
+ container arvados.Container,
+ status chan arvados.Container) {
log.Printf("Monitoring container %v started", container.UUID)
defer log.Printf("Monitoring container %v finished", container.UUID)
// Mutex between squeue sync and running sbatch or scancel.
squeueUpdater.SlurmLock.Lock()
- err := scancelCmd(container).Run()
+ cmd := scancelCmd(container)
+ msg, err := cmd.CombinedOutput()
squeueUpdater.SlurmLock.Unlock()
if err != nil {
- log.Printf("Error stopping container %s with scancel: %v",
- container.UUID, err)
+ log.Printf("Error stopping container %s with %v %v: %v %v",
+ container.UUID, cmd.Path, cmd.Args, err, string(msg))
if squeueUpdater.CheckSqueue(container.UUID) {
log.Printf("Container %s is still in squeue after scancel.",
container.UUID)
}
monitorDone = true
}
+
+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
+}