// If the container is marked as Locked, check if it is already in the slurm
// queue. If not, submit it.
//
// 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) {
submitted := false
for !*monitorDone {
if squeueUpdater.CheckSqueue(container.UUID) {
// Found in the queue, so continue monitoring
submitted = true
} else if container.State == dispatch.Locked && !submitted {
// Not in queue but in Locked state and we haven't
// submitted it yet, so submit it.
log.Printf("About to submit queued container %v", container.UUID)
if _, err := submit(dispatcher, container, *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)
}
submitted = true
} else {
// Not in queue and we are not going to submit it.
// Refresh the container state. If it is
// Complete/Cancelled, do nothing, if it is Locked then
// release it back to the Queue, if it is Running then
// clean up the record.
var con dispatch.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
case dispatch.Running:
st = dispatch.Cancelled
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)
}
}
}
func doMain() error {
flags := flag.NewFlagSet("crunch-dispatch-local", flag.ExitOnError)
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")
// Parse args; omit the first arg which is the command name
flags.Parse(os.Args[1:])
runningCmds = make(map[string]*exec.Cmd)
arv, err := arvadosclient.MakeArvadosClient()
if err != nil {
log.Printf("Error making Arvados client: %v", err)
return err
}
arv.Retries = 25
dispatcher := dispatch.Dispatcher{
Arv: arv,
RunContainer: run,
PollInterval: time.Duration(*pollInterval) * time.Second,
DoneProcessing: make(chan struct{})}
err = dispatcher.RunDispatcher()
if err != nil {
return err
}
runningCmdsMutex.Lock()
// Finished dispatching; interrupt any crunch jobs that are still running
for _, cmd := range runningCmds {
cmd.Process.Signal(os.Interrupt)
}
runningCmdsMutex.Unlock()
// Wait for all running crunch jobs to complete / terminate
waitGroup.Wait()
return nil
}
func testWithServerStub(c *C, apiStubResponses map[string]arvadostest.StubResponse, crunchCmd string, expected string) {
apiStubResponses["/arvados/v1/api_client_authorizations/current"] =
arvadostest.StubResponse{200, string(`{"uuid": "` + arvadostest.Dispatch1AuthUUID + `", "api_token": "xyz"}`)}
apiStub := arvadostest.ServerStub{apiStubResponses}
api := httptest.NewServer(&apiStub)
defer api.Close()
arv := arvadosclient.ArvadosClient{
Scheme: "http",
ApiServer: api.URL[7:],
ApiToken: "abc123",
Client: &http.Client{Transport: &http.Transport{}},
Retries: 0,
}
buf := bytes.NewBuffer(nil)
log.SetOutput(io.MultiWriter(buf, os.Stderr))
defer log.SetOutput(os.Stderr)
*crunchRunCommand = crunchCmd
doneProcessing := make(chan struct{})
dispatcher := dispatch.Dispatcher{
Arv: arv,
PollInterval: time.Duration(1) * time.Second,
RunContainer: func(dispatcher *dispatch.Dispatcher,
container dispatch.Container,
status chan dispatch.Container) {
run(dispatcher, container, status)
doneProcessing <- struct{}{}
},
DoneProcessing: doneProcessing}
startCmd = func(container dispatch.Container, cmd *exec.Cmd) error {
dispatcher.UpdateState(container.UUID, "Running")
dispatcher.UpdateState(container.UUID, "Complete")
return cmd.Start()
}
go func() {
for i := 0; i < 80 && !strings.Contains(buf.String(), expected); i++ {
time.Sleep(100 * time.Millisecond)
}
dispatcher.DoneProcessing <- struct{}{}
}()
err := dispatcher.RunDispatcher()
c.Assert(err, IsNil)
// Wait for all running crunch jobs to complete / terminate
waitGroup.Wait()
c.Check(buf.String(), Matches, `(?ms).*`+expected+`.*`)
}
func doMain() error {
flags := flag.NewFlagSet("crunch-dispatch-slurm", flag.ExitOnError)
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")
// Parse args; omit the first arg which is the command name
flags.Parse(os.Args[1:])
arv, err := arvadosclient.MakeArvadosClient()
if err != nil {
log.Printf("Error making Arvados client: %v", err)
return err
}
arv.Retries = 25
squeueUpdater.StartMonitor(time.Duration(*pollInterval) * time.Second)
defer squeueUpdater.Done()
dispatcher := dispatch.Dispatcher{
Arv: arv,
RunContainer: run,
PollInterval: time.Duration(*pollInterval) * time.Second,
DoneProcessing: make(chan struct{})}
err = dispatcher.RunDispatcher()
if err != nil {
return err
}
return nil
}
// Run or monitor a container.
//
// 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) {
log.Printf("Monitoring container %v started", container.UUID)
defer log.Printf("Monitoring container %v finished", container.UUID)
monitorDone := false
go monitorSubmitOrCancel(dispatcher, container, &monitorDone)
for container = range status {
if container.State == dispatch.Locked || container.State == dispatch.Running {
if container.Priority == 0 {
log.Printf("Canceling container %s", container.UUID)
// Mutex between squeue sync and running sbatch or scancel.
squeueUpdater.SlurmLock.Lock()
err := scancelCmd(container).Run()
squeueUpdater.SlurmLock.Unlock()
if err != nil {
log.Printf("Error stopping container %s with scancel: %v",
container.UUID, err)
if squeueUpdater.CheckSqueue(container.UUID) {
log.Printf("Container %s is still in squeue after scancel.",
container.UUID)
continue
}
}
err = dispatcher.UpdateState(container.UUID, dispatch.Cancelled)
}
}
}
monitorDone = true
}
func testWithServerStub(c *C, apiStubResponses map[string]arvadostest.StubResponse, crunchCmd string, expected string) {
apiStub := arvadostest.ServerStub{apiStubResponses}
api := httptest.NewServer(&apiStub)
defer api.Close()
arv := arvadosclient.ArvadosClient{
Scheme: "http",
ApiServer: api.URL[7:],
ApiToken: "abc123",
Client: &http.Client{Transport: &http.Transport{}},
Retries: 0,
}
buf := bytes.NewBuffer(nil)
log.SetOutput(io.MultiWriter(buf, os.Stderr))
defer log.SetOutput(os.Stderr)
crunchRunCommand = &crunchCmd
doneProcessing := make(chan struct{})
dispatcher := dispatch.Dispatcher{
Arv: arv,
PollInterval: time.Duration(1) * time.Second,
RunContainer: func(dispatcher *dispatch.Dispatcher,
container dispatch.Container,
status chan dispatch.Container) {
go func() {
time.Sleep(1 * time.Second)
dispatcher.UpdateState(container.UUID, dispatch.Running)
dispatcher.UpdateState(container.UUID, dispatch.Complete)
}()
run(dispatcher, container, status)
doneProcessing <- struct{}{}
},
DoneProcessing: doneProcessing}
go func() {
for i := 0; i < 80 && !strings.Contains(buf.String(), expected); i++ {
time.Sleep(100 * time.Millisecond)
}
dispatcher.DoneProcessing <- struct{}{}
}()
err := dispatcher.RunDispatcher()
c.Assert(err, IsNil)
c.Check(buf.String(), Matches, `(?ms).*`+expected+`.*`)
}
func (s *TestSuite) TestIntegration(c *C) {
arv, err := arvadosclient.MakeArvadosClient()
c.Assert(err, IsNil)
echo := "echo"
crunchRunCommand = &echo
doneProcessing := make(chan struct{})
dispatcher := dispatch.Dispatcher{
Arv: arv,
PollInterval: time.Duration(1) * time.Second,
RunContainer: func(dispatcher *dispatch.Dispatcher,
container dispatch.Container,
status chan dispatch.Container) {
run(dispatcher, container, status)
doneProcessing <- struct{}{}
},
DoneProcessing: doneProcessing}
startCmd = func(container dispatch.Container, cmd *exec.Cmd) error {
dispatcher.UpdateState(container.UUID, "Running")
dispatcher.UpdateState(container.UUID, "Complete")
return cmd.Start()
}
err = dispatcher.RunDispatcher()
c.Assert(err, IsNil)
// Wait for all running crunch jobs to complete / terminate
waitGroup.Wait()
// There should be no queued containers now
params := arvadosclient.Dict{
"filters": [][]string{[]string{"state", "=", "Queued"}},
}
var containers dispatch.ContainerList
err = arv.List("containers", params, &containers)
c.Check(err, IsNil)
c.Assert(len(containers.Items), Equals, 0)
// Previously "Queued" container should now be in "Complete" state
var container dispatch.Container
err = arv.Get("containers", "zzzzz-dz642-queuedcontainer", nil, &container)
c.Check(err, IsNil)
c.Check(container.State, Equals, "Complete")
}
// Run a container.
//
// If the container is Locked, start a new crunch-run process and wait until
// crunch-run completes. If the priority is set to zero, set an interrupt
// signal to the crunch-run process.
//
// If the container is in any other state, or is not Complete/Cancelled after
// crunch-run terminates, mark the container as Cancelled.
func run(dispatcher *dispatch.Dispatcher,
container dispatch.Container,
status chan dispatch.Container) {
uuid := container.UUID
if container.State == dispatch.Locked {
waitGroup.Add(1)
cmd := exec.Command(*crunchRunCommand, uuid)
cmd.Stdin = nil
cmd.Stderr = os.Stderr
cmd.Stdout = os.Stderr
log.Printf("Starting container %v", uuid)
// Add this crunch job to the list of runningCmds only if we
// succeed in starting crunch-run.
runningCmdsMutex.Lock()
if err := startCmd(container, cmd); err != nil {
runningCmdsMutex.Unlock()
log.Printf("Error starting %v for %v: %q", *crunchRunCommand, uuid, err)
dispatcher.UpdateState(uuid, dispatch.Cancelled)
} else {
runningCmds[uuid] = cmd
runningCmdsMutex.Unlock()
// Need to wait for crunch-run to exit
done := make(chan struct{})
go func() {
if _, err := cmd.Process.Wait(); err != nil {
log.Printf("Error while waiting for crunch job to finish for %v: %q", uuid, err)
}
log.Printf("sending done")
done <- struct{}{}
}()
Loop:
for {
select {
case <-done:
break Loop
case c := <-status:
// Interrupt the child process if priority changes to 0
if (c.State == dispatch.Locked || c.State == dispatch.Running) && c.Priority == 0 {
log.Printf("Sending SIGINT to pid %d to cancel container %v", cmd.Process.Pid, uuid)
cmd.Process.Signal(os.Interrupt)
}
}
}
close(done)
log.Printf("Finished container run for %v", uuid)
// Remove the crunch job from runningCmds
runningCmdsMutex.Lock()
delete(runningCmds, uuid)
runningCmdsMutex.Unlock()
}
waitGroup.Done()
}
// If the container is not finalized, then change it to "Cancelled".
err := dispatcher.Arv.Get("containers", uuid, nil, &container)
if err != nil {
log.Printf("Error getting final container state: %v", err)
}
if container.LockedByUUID == dispatcher.Auth.UUID &&
(container.State == dispatch.Locked || container.State == dispatch.Running) {
log.Printf("After %s process termination, container state for %v is %q. Updating it to %q",
*crunchRunCommand, container.State, uuid, dispatch.Cancelled)
dispatcher.UpdateState(uuid, dispatch.Cancelled)
}
// drain any subsequent status changes
for _ = range status {
}
log.Printf("Finalized container %v", uuid)
}
func (s *TestSuite) integrationTest(c *C,
newSqueueCmd func() *exec.Cmd,
sbatchCmdComps []string,
runContainer func(*dispatch.Dispatcher, dispatch.Container)) dispatch.Container {
arvadostest.ResetEnv()
arv, err := arvadosclient.MakeArvadosClient()
c.Assert(err, IsNil)
var sbatchCmdLine []string
// Override sbatchCmd
defer func(orig func(dispatch.Container) *exec.Cmd) {
sbatchCmd = orig
}(sbatchCmd)
sbatchCmd = func(container dispatch.Container) *exec.Cmd {
sbatchCmdLine = sbatchFunc(container).Args
return exec.Command("sh")
}
// Override squeueCmd
defer func(orig func() *exec.Cmd) {
squeueCmd = orig
}(squeueCmd)
squeueCmd = newSqueueCmd
// There should be no queued containers now
params := arvadosclient.Dict{
"filters": [][]string{[]string{"state", "=", "Queued"}},
}
var containers dispatch.ContainerList
err = arv.List("containers", params, &containers)
c.Check(err, IsNil)
c.Check(len(containers.Items), Equals, 1)
echo := "echo"
crunchRunCommand = &echo
doneProcessing := make(chan struct{})
dispatcher := dispatch.Dispatcher{
Arv: arv,
PollInterval: time.Duration(1) * time.Second,
RunContainer: func(dispatcher *dispatch.Dispatcher,
container dispatch.Container,
status chan dispatch.Container) {
go runContainer(dispatcher, container)
run(dispatcher, container, status)
doneProcessing <- struct{}{}
},
DoneProcessing: doneProcessing}
squeueUpdater.StartMonitor(time.Duration(500) * time.Millisecond)
err = dispatcher.RunDispatcher()
c.Assert(err, IsNil)
squeueUpdater.Done()
c.Check(sbatchCmdLine, DeepEquals, sbatchCmdComps)
// There should be no queued containers now
err = arv.List("containers", params, &containers)
c.Check(err, IsNil)
c.Check(len(containers.Items), Equals, 0)
// Previously "Queued" container should now be in "Complete" state
var container dispatch.Container
err = arv.Get("containers", "zzzzz-dz642-queuedcontainer", nil, &container)
c.Check(err, IsNil)
return container
}
