Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

New materials are updated #4

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
+77 −50
Open

New materials are updated #4

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
127 changes: 77 additions & 50 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,24 +8,65 @@ ospool:
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

## Diagnostics with condor_q
- [Job is held](#job-is-held)
- [Job completed but was unsuccessful](#job-completed-but-was-unsuccessful)
- [Job does not start](#Job-does-not-start)
- [Job is running longer than expected](#job-is-running-longer-than-expected)

The `condor_q` command shows the status of the jobs and it can be used
to diagnose why jobs are not running. Using the `-better-analyze` flag
with `condor_q` can show you detailed information about why a job isn't
starting on a specific pool. Since OSG Connect sends jobs to many places, we also need to
specify a pool name with the `-pool` flag.
## 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

Unless you know a specific pool you would like to query, checking the `flock.opensciencegrid.org` pool is usually a good place to start.
$ 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 ⋯

$ condor_q -better-analyze JOB-ID -pool POOL-NAME
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](https://www.google.com/url?q=https://htcondor.readthedocs.io/en/latest/man-pages/condor_chirp.html&sa=D&source=editors&ust=1740162417565289&usg=AOvVaw1HJ8p454psoN6alPpujatn) 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@login.osgconnect.net
$ ssh username@apxx...

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

Expand All @@ -34,10 +75,9 @@ 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. Remember that you will also need to
specify a pool name. In this case we'll use `flock.opensciencegrid.org`:
-better-analyze` to find out.

$ condor_q -better-analyze JOB-ID -pool flock.opensciencegrid.org
$ condor_q -better-analyze JOB-ID

# Produces a long ouput.
# The following lines are part of the output regarding the job requirements.
Expand All @@ -53,7 +93,7 @@ specify a pool name. In this case we'll use `flock.opensciencegrid.org`:
[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 51200 MB of memory.
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:

Expand Down Expand Up @@ -87,39 +127,26 @@ Alternatively, you can edit the resource requirements of the idle job in queue:

condor_qedit JOB_ID RequestMemory 2048


## Held jobs and condor_release

Occasionally, a job can fail in various ways and go into "Held"
state. Held state means that the job has encountered some error, and
cannot run. This doesn't necessarily mean that your job has failed, but,
for whatever reason, Condor cannot fulfill your request(s).

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

transfer_output_files = outputfile

However, when the job executed, it went into Held state:

$ condor_q -analyze 372993.0
-- Submitter: login01.osgconnect.net : <192.170.227.195:56174> : login01.osgconnect.net
---
372993.000: Request is held.
Hold reason: Error from [email protected]: STARTER at 10.3.11.39 failed to send file(s) to <192.170.227.195:40485>: error reading from /wntmp/condor/compute-6-28/execute/dir_9368/glide_J6I1HT/execute/dir_16393/outputfile: (errno 2) No such file or directory; SHADOW failed to receive file(s) from <192.84.86.100:50805>

Let's break down this error message piece by piece:

Hold reason: Error from [email protected]: STARTER at 10.3.11.39 failed to send file(s) to <192.170.227.195:40485>

This part is quite cryptic, but it simply means that the worker node
where your job executed ([email protected] or 10.3.11.39)
tried to transfer a file to the OSG Connect login node (192.170.227.195)
but did not succeed. The next part explains why:

error reading from /wntmp/condor/compute-6-28/execute/dir_9368/glide_J6I1HT/execute/dir_16393/outputfile: (errno 2) No such file or directory

This bit has the full path of the file that Condor tried to transfer back to `login.osgconnect.net`. 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`! Condor 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 is running longer than expected

To troubleshoot this issue we recommend checking your .log files and see if
<ul>
<li>The job has been continuously running on the same slot</li>
<ul>
Debugging tools:
<li>[condor_tail](https://htcondor.readthedocs.io/en/latest/man-pages/condor_tail.html)</li>- It returns the last X bytes of the job
<li>[condor_ssh_to_job](https://htcondor.readthedocs.io/en/latest/man-pages/condor_ssh_to_job.html). It allows to log in to the execution point. Point to be noted-It **does not work** on all systems.
</ul>
<li>The job has been interrupted and restarted on another slot</li>
<ul>
Debugging tips:
<li>If it happens once or twice, adjust your expectation of the runtime.</li>
<li>If it happens many times, your job runtime may be too long for the system or [contact the support team](maito:[email protected]) expectation of the runtime.</li>
</ul>
<li>The job is stuck on the file transfer step</li>
<ul>
Debugging tips:
<li>Check the `.log` file for meaning errors</li>
<li> Can happen if you or someone else is transferring a lot of data (large size or many files) and the `Access Point` is overwhelmed</li>
</ul>
</ul>