Monday Exercise 3.3: Retries

The goal of this exercise is to demonstrate running a job that intermittently fails and thus could benefit from having HTCondor automatically retry it.

This first part of the exercise should take only a few minutes, and is designed to setup the next exercises.

Bad Job

Let’s assume that a colleague has shared with you a program, and it fails once in a while. In the real world, we would probably just fix the program, but what if you cannot change the software? Unfortunately, this situation happens more often than we would like.

Below is a simple Python script that fails once in a while. We will not fix it, but use it to simulate a program that can fail and that we cannot fix.

#!/usr/bin/env python

# murphy.py simulates a real program with real problems
import random
import sys
import time

# Create a random number seeded by system entropy
r = random.SystemRandom()

# One time in three, simulate a runtime error
if (r.randint(0,2) == 0):
    # intentionally print no output
    sys.exit(15)
else:
    time.sleep(3)
    print "All work done correctly"

# By convention, zero exit code means success
sys.exit(0)

Even if you are not a Python expert, you may be able to figure out what this program does.

Let’s see what happens when a program like this one is run in HTCondor.

  1. In a new directory for this exercise, save the script above as murphy.py.
  2. Write a submit file for the script; queue 20 instances of the job and be sure to ask for 20 MB of memory and disk.
  3. Submit the file and wait for the jobs to finish.

What output do you expect? What output did you get? If you are curious about the exit code from the job, it is saved in completed jobs in condor_history in the ExitCode attribute. The following command will show the ExitCode for a given cluster of jobs:

[email protected] $ condor_history <CLUSTER> -af ProcId ExitCode

(Be sure to replace <cluster> with your actual cluster ID)

How many of the jobs succeeded? How many failed?

Retrying Failed Jobs

Now let’s see if we can solve the problem of jobs that fail once in a while. In this particular case, if HTCondor runs a failed job again, it has a good chance of succeeding. Not all failing jobs are like this, but in this case it is a reasonable assumption.

From the lecture materials, implement the max_retries feature to retry any job with a non-zero exit code up to 5 times, then resubmit the jobs. Did your change work?

After the jobs have finished, examine the log file(s) to see what happened in detail. Did any jobs need to be restarted? Another way to see how many restarts there were is to look at the NumJobStarts attribute of a completed job with the condor_history command, in the same way you looked at the ExitCode attribute earlier. Does the number of retries seem correct? For those jobs which did need to be retried, what is their ExitCode; and what about the ExitCode from earlier execution attempts?

A (Too) Long Running Job

Sometimes, an ill-behaved job will get stuck in a loop and run forever, instead of exiting with a failure code, and it may just need to be re-run (or run on a different execute server) to complete without getting stuck. We can modify our Python program to simulate this kind of bad job with the following file:

#!/usr/bin/env python

# murphy.py simulate a real program with real problems
import random
import sys
import time

# Create a random number seeded by system entropy
r = random.SystemRandom()

# One time in three, simulate an infinite loop
if (r.randint(0,2) == 0):
        # intentionally print no output
        time.sleep(3600)
        sys.exit(15)
else:
        time.sleep(3)
        print "All work done correctly"

# By convention, zero exit code means success
sys.exit(0)

Again, you may be able to figure out what this new program does.

  1. Save the script to a new file named murphy2.py.
  2. Copy your previous submit file to a new name and change the executable to murphy2.py.
  3. If you like, submit the new file — but after a while be sure to remove the whole cluster to clear out the “hung” jobs.
  4. Now try to change the submit file to automatically remove any jobs that run for more than one minute. You can make this change with just a single line in your submit file

    periodic_remove = (JobStatus == 2) && ( (CurrentTime - EnteredCurrentStatus) > 60 )
    
  5. Submit the new file. Do the long running jobs get removed? What does condor_history show for the cluster after all jobs are done? Which job status (i.e. idle, held, running) do you think JobStatus == 2 corresponds to?

Bonus Exercise

If you have time, edit your submit file so that instead of removing long running jobs, have HTCondor automatically put the long-running job on hold, and then automatically release it.