Skip to content

Latest commit

 

History

History
152 lines (108 loc) · 6.85 KB

File metadata and controls

152 lines (108 loc) · 6.85 KB
ospool
path
htc_workloads/submitting_workloads/tutorial-error101/README.md

Troubleshooting Job Errors

In this lesson, we'll learn how to troubleshoot jobs that never start or fail in unexpected ways.

Troubleshooting techniques

We will look into 4 criterias. The criteria's are listed below

Job is held

HTCondor puts your jobs on hold when something goes wrong in the process of managing your jobs. HTCondor provides a Hold Reason that explains what went wrong. To see what went wrong you can use the command condor_q -hold. A typical hold message may look like the following

$ condor_q -hold
ap40.uw.osg-htc.org : <128.105.68.62:9618?... @ 08/01/24 15:01:50
ID         OWNER          HELD_SINCE  HOLD_REASON
130.0       alice           8/1  14:56 Transfer input files failure at ⋯

In this particular case, a user had this in his or her HTCondor submit file:

transfer_output_files = outputfile

However, when the job executed, it went into Held state. To see more about the error message condor_q -hold JOB-ID can be used. An example error message in this scenario is shown below:

$ condor_q -hold JOB-ID

Transfer output files failure at access point… while receiving files from the execution point. Details: Error from ….execute point … failed to send file(s) at apxx; failed to read from file /path: (errno 2) No such file or directory

The reason why the file transfer failed is because outputfile was never created on the worker node. Remember that at the beginning we said that the user specifically requested transfer_outputfiles = outputfile! HTCondor could not complete this request, and so the job went into Held state instead of finishing normally.

It's quite possible that the error was simply transient, and if we retry, the job will succeed. We can re-queue a job that is in Held state by using condor_release:

condor_release JOB-ID 

Job completed but was unsuccessful

Under this case two things will be considered. First, your job has run but the code did not execute correctly or in the expected manner. Second, your job ran but it did not produce/transferred the desired output files.

For these secnarios having various ls or echo statemets that shows the contents of the directory and code's progression is helpful in diagnosing the actual issue. An example is shown here:

echo "Here are the files at the start of the job:"
ls -R
echo "Executing main command..."
my_command	#Part of the original code
echo "... finished. Here are the files at the end of the job:"
ls -R

For your commands if eanbling debugging or verbose logging is an option then do those. Moreover, condor_chirp can be a very useful tool in this regard as it sends information directly back to the Access Point. For example, the following will add a statement to your .log file: command1 condor_chirp ulog "Finished executing command1" command2

Job does not start

Matchmaking cycle can take more than 5 minutes to complete and it can be longer if the server is busy. Specially, the more/more-specific resources are requested, the longer is the wait. For example if you are submitting a lot of GPU jobs, there may not be enough GPUs in the pool. This type of issues can be diagnosed using the -better-analyze flag with condor_q to see the detailed information about why a job isn't starting.

$ condor_q -better-analyze JOB-ID 

Let's do an example. First we'll need to login as usual, and then load the tutorial error101.

$ ssh username@apxx...

$ git clone https://github.com/OSGConnect/tutorial-error101.git
$ cd tutorial-error101
$ condor_submit error101_job.submit 

We'll check the job status the normal way:

condor_q username

For some reason, our job is still idle. Why? Try using condor_q -better-analyze to find out.

$ condor_q -better-analyze JOB-ID 
 
# Produces a long ouput. 
# The following lines are part of the output regarding the job requirements.  

The Requirements expression for your job reduces to these conditions:

         Slots
Step    Matched  Condition
-----  --------  ---------
[0]       10674  TARGET.Arch == "X86_64"
[1]       10674  TARGET.OpSys == "LINUX"
[3]       10674  TARGET.Disk >= RequestDisk
[5]           0  TARGET.Memory >= RequestMemory
[8]       10674  TARGET.HasFileTransfer

By looking through the match conditions, we see that many nodes match our requests for the Linux operating system and the x86_64 architecture, but none of them match our requirement for memory.

Let's look at our submit script and see if we can find the source of this error:

$ cat error101_job.submit 
Universe = vanilla

Executable = error101.sh

# to sleep an hour
Arguments = 3600

request_memory = 2 TB

Error = job.err 
Output = job.out 
Log = job.log 
Queue 1 

See the request_memory line? We are asking for 2 Terabytes of memory, when we meant to only ask for 2 Gigabytes of memory. Our job is not matching any available job slots because none of the slots offer 2 TB of memory. Let's fix that by changing that line to read request_memory = 2 GB.

$ nano error101_job.submit

Let's cancel our idle job with the condor_rm command and then resubmit our edited job:

$ condor_rm JOB-ID
$ condor_submit error101_job.submit

Alternatively, you can edit the resource requirements of the idle job in queue:

condor_qedit JOB_ID RequestMemory 2048

Job is running longer than expected

To troubleshoot this issue we recommend checking your .log files and see if

  • The job has been continuously running on the same slot
  • The job has been interrupted and restarted on another slot
    • Debugging tips:
    • If it happens once or twice, adjust your expectation of the runtime.
    • If it happens many times, your job runtime may be too long for the system or [contact the support team](maito:support@osg-htc.org) expectation of the runtime.
  • The job is stuck on the file transfer step
    • Debugging tips:
    • Check the `.log` file for meaning errors
    • Can happen if you or someone else is transferring a lot of data (large size or many files) and the `Access Point` is overwhelmed