Starting and Stopping a Session

The following code segment shows the most basic DRMAA python binding program

example1:

 1  #!/usr/bin/env python
 2
 3  import DRMAA
 4
 5  def main():
 6      """Create a DRMAA session and exit"""
 7      s=DRMAA.Session()
 8      print 'A DRMAA object was created'
 9      s.init()
10      print 'A session was started successfully'
11      s.exit()
12
13  if __name__=='__main__':
14      main()

The first thing to notice is that every call to a DRMAA function will return an error code. In this tutorial, we ignore all error codes.

Now let's look at the functions being called. First, on line 7, we initialise a Session object by calling DRMAA.Session(). Then we initialise the session in line 9 by init() This function sets up the DRMAA session and must be called before most other DRMAA functions. Some functions, like getContact(), can be called before init(), but these functions only provide general information. Any function that does work, such as runJob() or wait() must be called after init() returns. If such a function is called before init() returns, it will return an error.

init() creates a session and starts an event client listener thread. The session is used for organizing jobs submitted through DRMAA, and the thread is used to receive updates from the queue master about the state of jobs and the system in general. Once init() has been called successfully, it is the responsibility of the calling application to also call exit() before terminating. If an application does not call exit() before terminating, session state may be left behind in the user's home directory (under .sge/drmaa), and the queue master may be left with a dead event client handle, which can decrease queue master performance.

At the end of our program, on line 11, we call exit(). exit() cleans up the session and stops the event client listener thread. Most other DRMAA functions must be called before exit(). Some functions, like getContact(), can be called after exit(), but these functions only provide general information. Any function that does work, such as runJob() or wait() must be called before exit() is called. If such a function is called after exit() is called, it will return an error.

example1.1:

 1  #!/usr/bin/env python
 2
 3  import DRMAA
 4
 5  def main():
 6      """Create a session, show that each session has an id,
 7      use session id to disconnect, then reconnect. Then exit"""
 8      s=DRMAA.Session()
 9      s.init()
10      print 'A session was started successfullly'
11      response = s.getContact()
12      print 'getContact() returns: ' + response
13      s.exit()
14      print 'Exited from session'
15
16      s.init(response)
17      print 'Session was restarted successfullly'
18      s.exit()
19
20
21  if __name__=='__main__':
22      main()

This example is very similar to Example 1. The difference is that it uses the Grid Engine feature of reconnectable sessions. The DRMAA concept of a session is translated into a session tag in the Grid Engine job structure. That means that every job knows to which session it belongs. With reconnectable sessions, it's possible to initialize the DRMAA library to a previous session, allowing the library access to that session's job list. The only limitation, though, is that jobs which end between the calls to exit() and init() will be lost, as the reconnecting session will no longer see these jobs, and so won't know about them.

Through line 10, this example is very similar to Example 1. On line 11, however, we use the getContact() function to get the contact information for this session. On line 13 we then exit the session. On line 16, we use the stored contact information to reconnect to the previous session. Had we submitted jobs before calling exit(), those jobs would now be available again for operations such as wait() and synchronize(). Finally, on line 18 we exit the session a second time.

Running a Job

The following code segment shows how to use the DRMAA python binding to submit a job to Grid Engine:

example2:

 1  #!/usr/bin/env python
 2
 3  import DRMAA
 4  import os
 5
 6  def main():
 7      """Submit a job.
 8      Note, need file called sleeper.sh in home directory. An example:
 9      echo 'Hello World $1'
10      """
11      s=DRMAA.Session()
12      s.init()
13
14      print 'Creating job template'
15      jt = s.createJobTemplate()
16      jt.remoteCommand = os.getcwd() + '/sleeper.sh'
17      jt.args = ['42','Simon says:']
18      jt.joinFiles=True
19      jt.outputPath=":"+DRMAA.JobTemplate.HOME_DIRECTORY+'/tmp/DRMAA_JOB_OUT'
20
21      jobid = s.runJob(jt)
22      print 'Your job has been submitted with id ' + jobid
23
24      print 'Cleaning up'
25      s.deleteJobTemplate(jt)
26      s.exit()
27
28  if __name__=='__main__':
29      main()

The beginning and end of this program are the same as the previous one. What's different is in lines 15-25. On line 15 we ask DRMAA to allocate a job template for us. A job template is a structure used to store information about a job to be submitted. The same template can be reused for multiple calls to runJob() or runBulkJob().

On line 16 we set the REMOTE_COMMAND attribute. This attribute tells DRMAA where to find the program we want to run. Its value is the path to the executable. The path be be either relative or absolute. If relative, it is relative to the WD attribute, which if not set defaults to the user's home directory. For more information on DRMAA attributes, please see the attributes man page. Note that for this program to work, the script "sleeper.sh" must be in the current directory.

On line 17 we set the V_ARGV attribute. This attribute tells DRMAA what arguments to pass to the executable. For more information on DRMAA attributes, please see the attributes man page.

On line 21 we submit the job with runJob(). DRMAA will place the id assigned to the job into the character array we passed to runJob(). The job is now running as though submitted by qsub. At this point calling exit() and/or terminating the program will have no effect on the job.

To clean things up, we delete the job template on line 25. This frees the memory DRMAA set aside for the job template, but has no effect on submitted jobs. Finally, on line 26, we call exit().

If instead of a single job we had wanted to submit an array job, we could have replaced the else on line 21 and 22 with the following:

example2.1:

21      jobid = s.runBulkJobs(jt,1,30,2)
22      print 'Your job has been submitted with id ' + str(jobid)

This code segment submits an array job with 15 tasks numbered 1, 3, 5, 7, etc. An important difference to note is that runBulkJobs() returns the job ids in an array. On line 22, we print all the job ids.

Waiting for a Job

Now we're going to extend our example to include waiting for a job to finish.

example3:

 1  #!/usr/bin/env python
 2
 3  import DRMAA
 4  import os
 5
 6  def main():
 7      """Submit a job and wait for it to finish.
 8      Note, need file called sleeper.sh in home directory. An example:
 9      echo 'Hello World $1'
10      """
11      s=DRMAA.Session()
12      s.init()
13
14      print 'Creating job template'
15      jt = s.createJobTemplate()
16      jt.remoteCommand = os.getcwd() + '/sleeper.sh'
17      jt.args = ['42','Simon says:']
18      jt.joinFiles=True
19      jt.outputPath=":"+DRMAA.JobTemplate.HOME_DIRECTORY+'/tmp/DRMAA_JOB_OUT'
20
21      jobid = s.runJob(jt)
22      print 'Your job has been submitted with id ' + jobid
23
24      retval = s.wait(jobid, DRMAA.Session.TIMEOUT_WAIT_FOREVER)
25      print 'Job: ' + str(retval.getJobId()) + ' finished with code ' + str(retval.getExitStatus())
26
27      print 'Cleaning up'
28      s.deleteJobTemplate(jt)
29      s.exit()
30
31  if __name__=='__main__':
32      main()

This example is very similar to Example 2 except for line 24. On line 24 we call wait() to wait for the job to end. We have to give wait() both the id of the job for which we want to wait and a place to write the id of the job for which we actually waited because the job id we pass in could be JOB_IDS_SESSION_ANY, in which case wait() must have a way of tell us which job is the one that made it return. We also have to pass to wait() how long we are willing to wait for the job to finish. This could be a number of seconds, or it could be either TIMEOUT_WAIT_FOREVER or TIMEOUT_NO_WAIT. Lastly, we also have to pass in a place to write the exit status and the usage information. The exit status is an opaque number that is passed to the w...() functions to get information about how the job exited. The usage information is a list of name=value pairs in a DRMAA values structure. The values structure works exactly the same as the ids structure we talked about in Example 2.1. Assuming the wait worked, we query the job's exit status on lines 25.

An alternative to wait() when working with multiple jobs, such as jobs submitted by runBulkJobs() or multiple calls to runJob() is synchronize(). synchronize() waits for a set of jobs to finish. To use synchronize(), we could replace lines 21-24 with the following:

example3.1:

21      joblist = s.runBulkJobs(jt,1,30,2)
22      print 'Your job has been submitted with id ' + str(joblist)
23
24      s.synchronize(joblist, DRMAA.Session.TIMEOUT_WAIT_FOREVER, True)

Line 21 now call runBulkJobs() so that we have several jobs for which to wait. On line 56, instead of calling wait(), we call synchronize(). synchronize() takes only three iteresting parameters. The first is the list of ids for which to wait. This list must be a NULL-terminated array of strings. If the special id, JOB_IDS_SESSION_ALL, appears in the array, synchronize() will wait for all jobs submitted via DRMAA during this session, i.e. since init() was called. The second is how long to wait for all the jobs in the list to finish. This is the same as the timeout parameter for wait(). The third is whether this call to synchronize() should clean up after the job. After a job completes, it leaves behind accounting information, such as exist status and usage, until either wait() or synchronize() with dispose set to true is called. It is the responsibility of the application to make sure one of these two functions is called for every job. Not doing so creates a memory leak. Note that calling synchronize() with dispose set to true flushes all accounting information for all jobs in the list. If you want to use synchronize() and still recover the accounting information, set dispose to false and call wait() for each job. To do this in Example 3, we would replace lines 21-24 with the following:

example3.2:

26      s.synchronize(joblist, DRMAA.Session.TIMEOUT_WAIT_FOREVER,False)
27      for curjob in joblist:
28          print 'Collecting job ' + curjob
29          retval = s.wait(curjob, DRMAA.Session.TIMEOUT_WAIT_FOREVER)
30          print 'Job: ' + str(retval.getJobId()) + ' finished with code ' + str(retval.getExitStatus())

What's different is that on line 26, we set dispose to false, and then on lines 27-30 we wait once for each job, printing the exit status and usage information as we did in Example 3. We pass JOB_IDS_SESSION_ANY to wait() as the job id because we already know that all the jobs have finished, so we don't really care in what order we process them. In an interactive system where we couldn't guarantee that more jobs wouldn't be submitted between the synchronize and the wait, we would have to store the job ids from the runBulkJobs() in an array and then wait for each job specifically. Otherwise, the wait() could end up waiting for a job submitted after the call to synchronize().

Controling a Job

Now let's look at an example of how to control a job from DRMAA:

example4:

 1  #!/usr/bin/env python
 2
 3  import DRMAA
 4  import os
 5
 6  def main():
 7      """Submit a job, then kill it.
 8      Note, need file called sleeper.sh in home directory. An example:
 9      echo 'Hello World $1'
10      """
11      s=DRMAA.Session()
12      s.init()
13
14      print 'Creating job template'
15      jt = s.createJobTemplate()
16      jt.remoteCommand = os.getcwd() + '/sleeper.sh'
17      jt.args = ['42','Simon says:']
18      jt.joinFiles=True
19      jt.outputPath=":"+DRMAA.JobTemplate.HOME_DIRECTORY+'/tmp/JOB_OUT'
20
21      jobid = s.runJob(jt)
22      print 'Your job has been submitted with id ' + jobid
23      # options are: SUSPEND, RESUME, HOLD, RELEASE, TERMINATE
24      s.control(jobid,DRMAA.Session.TERMINATE)
25
26      print 'Cleaning up'
27      s.deleteJobTemplate(jt)
28      s.exit()
29
30  if __name__=='__main__':
31      main()

This example is very similar to Example 2 except for line 24. On line 24 we use control() to delete the job we just submitted. Aside from deleting the job, we could have also used control() to suspend, resume, hold, or release it. For more information, see the control man page.

Note that control() can be used to control jobs not submitted through DRMAA. Any valid SGE job id could be passed to control() as the id of the job to delete.

Getting Job Status

Here's an example of using DRMAA to query the status of a job:

example5:

 1  #!/usr/bin/env python
 2
 3  import DRMAA
 4  import time
 5  import os
 6
 7  def main():
 8      """Submit a job, and check its progress.
 9      Note, need file called sleeper.sh in home directory. An example:
10      echo 'Hello World $1'
11      sleep 30s
12      """
13      s=DRMAA.Session()
14      s.init()
15
16      print 'Creating job template'
17      jt = s.createJobTemplate()
18      jt.remoteCommand = os.getcwd() + '/sleeper.sh'
19      jt.args = ['42','Simon says:']
20      jt.joinFiles=True
21      jt.outputPath=":"+DRMAA.JobTemplate.HOME_DIRECTORY+'/tmp/JOB_OUT'
22
23      jobid = s.runJob(jt)
24      print 'Your job has been submitted with id ' + jobid
25
26      # Who needs a case statement when you have dictionaries?
27      decodestatus = {
28          DRMAA.Session.UNDETERMINED: 'process status cannot be determined',
29          DRMAA.Session.QUEUED_ACTIVE: 'job is queued and active',
30          DRMAA.Session.SYSTEM_ON_HOLD: 'job is queued and in system hold',
31          DRMAA.Session.USER_ON_HOLD: 'job is queued and in user hold',
32          DRMAA.Session.USER_SYSTEM_ON_HOLD: 'job is queued and in user and system hold',
33          DRMAA.Session.RUNNING: 'job is running',
34          DRMAA.Session.SYSTEM_SUSPENDED: 'job is system suspended',
35          DRMAA.Session.USER_SUSPENDED: 'job is user suspended',
36          DRMAA.Session.DONE: 'job finished normally',
37          DRMAA.Session.FAILED: 'job finished, but failed',
38          }
39
40      for ix in range(10):
41          print 'Checking ' + str(ix) + ' of 10 times'
42          status = s.getJobProgramStatus(jobid)
43          print decodestatus.get(status)
44          time.sleep(5)
45
46      print 'Cleaning up'
47      s.deleteJobTemplate(jt)
48      s.exit()
49
50  if __name__=='__main__':
51      main()

Again, this example is very similar to Example 2, this time with the exception of lines 26-44. First, after submitting the job, we sleep for 20 seconds to give SGE time to schedule the job. Then, on line 42, we use getJobProgramStatus() to get the status of the job. Line 43 determine what the job status is and report it.

Getting DRM information

Lastly, let's look at how to query the DRMAA library for information about the DRMS and the DRMAA implementation itself:

example6:

 1  #!/usr/bin/env python
 2
 3  import DRMAA
 4
 5  def main():
 6      """Query the system."""
 7      s=DRMAA.Session()
 8      print 'A DRMAA object was created'
 9      print 'Supported contact strings: ' + s.getContact()
10      print 'Supported DRM systems: ' + str(s.getDRMSInfo())
11      print 'Supported DRMAA implementations: ' + str(s.getDRMAAImplementation())
12
13
14      s.init()
15      print 'A session was started successfully'
16      print 'Supported contact strings: ' + s.getContact()
17      print 'Supported DRM systems: ' + str(s.getDRMSInfo())
18      print 'Supported DRMAA implementations: ' + str(s.getDRMAAImplementation())
19      print 'Version ' + str(s.getVersion())
20
21      print 'Exiting'
22      s.exit()
23
24
25  if __name__=='__main__':
26      main()

On line 9, we get the contact string list. This is the list of contact strings that will be understood by this DRMAA instance. Normally on of these strings is used to select to which DRM this DRMAA instance should be bound. In the Grid Engine 6.0 implementation, the contact string list is empty because there is only ever one possible DRM to which to bind. On line 10, we get the list of supported DRM systems. For the Grid Engine 6.0 implementation, this will always be Grid Engine 6.0. On line 11, we get the list of supported DRMAA implementations. For the Grid Engine 6.0 implementation, this will always be Grid Engine 6.0. On line 14, we call init(). After init() has been called, the get_contact() and get_DRM_system() calls change. On line 16, we call get_contact() again, this time to get the contact string that was used to bind to a DRMS in init(). For the Grid Engine 6.0 implementation, this will always be an empty string. On line 17, we call getDRMSInfo() again, this time to get the name of the DRMS to which DRMAA is bound. For the Grid Engine 6.0 implementation, this will always be Grid Engine 6.0. On line 18, we call get_implementation() again, this time to get the name of the DRMAA implementation to which the application is bound. For the Grid Engine 6.0 implementation, this will always be Grid Engine 6.0. On line 19, we get the version number of the DRMAA C binding specification supported by this DRMAA implementation. For the Grid Engine 6.0 implementation this is currently version 0.8. Finally, on line 22, we end the session with exit().