Self-Checkpointing Applications

This section is about writing jobs for an executable which periodically saves checkpoint information, and how to make HTCondor store that information safely, in case it’s needed to continue the job on another machine or at a later time.

This section is not about how to checkpoint a given executable; that’s up to you or your software provider.

How To Run Self-Checkpointing Jobs

The best way to run self-checkpointing code is to set checkpoint_exit_code in your submit file. (Any exit code will work, but if you can choose, consider error code 85. On Linux systems, this is ERESTART, which seems appropriate.) If the executable exits with checkpoint_exit_code, HTCondor will transfer the checkpoint to the submit node, and then immediately restart the executable in the same sandbox on the same machine, with the same arguments. This immediate transfer makes the checkpoint available for continuing the job even if the job is interrupted in a way that doesn’t allow for files to be transferred (e.g., power failure), or if the file transfer doesn’t complete in the time allowed.

For a job to use checkpoint_exit_code successfully, its executable must meet a number of requirements.

Requirements

Your self-checkpointing code may not meet all of the following requirements. In many cases, however, you will be able to add a wrapper script, or modify an existing one, to meet these requirements. (Thus, your executable may be a script, rather than the code that’s writing the checkpoint.) If you can not, consult Working Around the Assumptions and/or the Other Options.

  1. Your executable exits after taking a checkpoint with an exit code it does not otherwise use.

    • If your executable does not exit when it takes a checkpoint, HTCondor will not transfer its checkpoint. If your executable exits normally when it takes a checkpoint, HTCondor will not be able to tell the difference between taking a checkpoint and actually finishing; that is, if the checkpoint code and the terminal exit code are the same, your job will never finish.

  2. When restarted, your executable determines on its own if a checkpoint is available, and if so, uses it.

    • If your job does not look for a checkpoint each time it starts up, it will start from scratch each time; HTCondor does not run a different command line when restarting a job which has taken a checkpoint.

  3. Starting your executable up from a checkpoint is relatively quick.

    • If starting your executable up from a checkpoint is relatively slow, your job may not run efficiently enough to be useful, depending on the frequency of checkpoints and interruptions.

Using checkpoint_exit_code

The following Python script (example.py) is a toy example of code that checkpoints itself. It counts from 0 to 10 (exclusive), sleeping for 10 seconds at each step. It writes a checkpoint file (containing the next number) after each nap, and exits with code 85 at count 3, 6, and 9. It exits with code 0 when complete.

#!/usr/bin/env python

import sys
import time

value = 0
try:
    with open('example.checkpoint', 'r') as f:
        value = int(f.read())
except IOError:
    pass

print("Starting from {0}".format(value))
for i in range(value,10):
    print("Computing timestamp {0}".format(value))
    time.sleep(10)
    value += 1
    with open('example.checkpoint', 'w') as f:
        f.write("{0}".format(value))
    if value%3 == 0:
        sys.exit(85)

print("Computation complete")
sys.exit(0)

The following submit file (example.submit) commands HTCondor to transfer the file example.checkpoint to the submit node whenever the script exits with code 85. If interrupted, the job will resume from the most recent of those checkpoints. Before version 8.9.8, you must include your checkpoint file(s) in transfer_output_files; otherwise HTCondor will not transfer it (them). Starting with version 8.9.8, you may instead use transfer_checkpoint_files, as documented on the condor_submit man page.

checkpoint_exit_code        = 85
transfer_output_files       = example.checkpoint
should_transfer_files       = yes

executable                  = example.py
arguments                   =

output                      = example.out
error                       = example.err
log                         = example.log

request_cpus            = 1
request_memory          = 512M
request_disk            = 1G

queue 1

This example does not remove the “checkpoint file” generated for timestep 9 when the executable completes. This could be done in example.py immediately before it exits, but that would cause the final file transfer to fail, if you specified the file in transfer_output_files. The script could instead remove the file and then re-create it empty, it desired.

How Frequently to Checkpoint

Obviously, the longer the code spends writing checkpoints, and the longer your job spends transferring them, the longer it will take for you to get the job’s results. Conversely, the more frequently the job transfers new checkpoints, the less time the job loses if it’s interrupted. For most users and for most jobs, taking a checkpoint about once an hour works well, and it’s not a bad duration to start experimenting with. A number of factors will skew this interval up or down:

  • If your job(s) usually run on resources with strict time limits, you may want to adjust how often your job checkpoints to minimize wasted time. For instance, if your job writes a checkpoint after each hour, and each checkpoint takes five minutes to write out and then transfer, your fifth checkpoint will finish twenty-five minutes into the fifth hour, and you won’t gain any benefit from the next thirty-five minutes of computation. If you instead write a checkpoint every eighty-four minutes, your job will only waste four minutes.

  • If a particular code writes larger checkpoints, or writes smaller checkpoints unusually slowly, you may want to take a checkpoint less frequently than you would for other jobs of a similar length, to keep the total overhead (delay) the same. The opposite is also true: if the job can take checkpoints particularly quickly, or the checkpoints are particularly small, the job could checkpoint more often for the same amount of overhead.

  • Some code naturally checkpoints at longer or shorter intervals. If a code writes a checkpoint every five minutes, it may make sense for the executable to wait for the code to write ten or more checkpoints before exiting (which asks HTCondor to transfer the checkpoint file(s)). If a job is a sequence of steps, the natural (or only possible) checkpoint interval may be between steps.

  • How long it takes to restart from a checkpoint. It should never take longer to restart from a checkpoint than to recompute from the beginning, but the restart process is part of the overhead of taking a checkpoint. The longer a code takes to restart, the less often the executable should exit.

Measuring how long it takes to make checkpoints is left as an exercise for the reader. Since version 8.9.1, however, HTCondor will report in the job’s log (if a log is enabled for that job) how long file transfers, including checkpoint transfers, took.

Debugging Self-Checkpointing Jobs

Because a job may be interrupted at any time, it’s valid to interrupt the job at any time and see if a valid checkpoint is transferred. To do so, use condor_vacate_job to evict the job. When that’s done (watch the user log), use condor_hold to put it on hold, so that it can’t restart while you’re looking at the checkpoint (and potentially, overwrite it). Finally, to obtain the checkpoint file(s) themselves, use the somewhat mis-named condor_evicted_files to ask where they are.

For example, if your job is ID 635.0, and is logging to the file job.log, you can copy the files in the checkpoint to a subdirectory of the current as follows:

$ condor_vacate_job 635.0

Wait for the job to finish being evicted; hit CTRL-C when you see ‘Job was evicted.’ and immediately hold the job.

$ tail --follow job.log
$ condor_hold 635.0

Copy the checkpoint files from the spool. Note that _condor_stderr and _condor_stdout are the files corresponding to the job’s output and error submit commands; they aren’t named correctly until the the job finishes.

$ condor_evicted_files get 635.0
Copied to '635.0'.
$ cd 635.0

Now examine the checkpoint files to see if they look right. When you’re done, release the job to see if it actually works right.

$ condor_release 635.0
$ condor_ssh_to_job 635.0

You may also want to remove your copy of checkpoint files:

$ cd ..; rm -fr 635.0

Working Around the Assumptions

The basic technique here is to write a wrapper script (or modify an existing one), so that the executable has the necessary behavior, even if the code does not.

  1. Your executable exits after taking a checkpoint with an exit code it does not otherwise use.

    • If your code exits when it takes a checkpoint, but not with a unique code, your wrapper script will have to determine, when the executable exits, if it did so because it took a checkpoint. If so, the wrapper script will have to exit with a unique code. If the code could usefully exit with any code, and the wrapper script therefore can not exit with a unique code, you can instead instruct HTCondor to consider being killed by a particular signal as a sign of successful checkpoint; set +SuccessCheckpointExitBySignal to TRUE and +SuccessCheckpointExitSignal to the particular signal. (If you do not set checkpoint_exit_code, you must set +WantFTOnCheckpoint.)

    • If your code does not exit when it takes a checkpoint, the wrapper script will have to determine when a checkpoint has been made, kill the program, and then exit with a unique code.

  2. When restarted, your executable determines on its own if a checkpoint is available, and if so, uses it.

    • If your code requires different arguments to start from a checkpoint, the wrapper script must check for the presence of a checkpoint and start the executable with correspondingly modified arguments.

  3. Starting your executable up from a checkpoint is relatively quick.

    • The longer the start-up delay, the slower the job’s overall progress. If your job’s progress is too slow as a result of start-up delay, and your code can take checkpoints without exiting, read the ‘Delayed Transfers’ and ‘Manual Transfers’ sections below.

Other Options

The preceding sections of this HOWTO explain how a job meeting the requirements can take checkpoints at arbitrary intervals and transfer them back to the submit node. Although this is the method of operation most likely to result in an interrupted job continuing from a valid checkpoint, other, less reliable options exist.

Delayed Transfers

This method is risky, because it does not allow your job to recover from any failure mode other than an eviction (and sometimes not even then). It may also require changes to your executable. The advantage of this method is that it doesn’t require your code to restart, or even a recent version of HTCondor.

The basic idea is to take checkpoints as the job runs, but not transfer them back to the submit node until the job is evicted. This implies that your executable doesn’t exit until the job is complete (which is the normal case). If your code has long start-up delays, you’ll naturally not want it to exit after it writes a checkpoint; otherwise, the wrapper script could restart the code as necessary.

To use this method, set when_to_transfer_output to ON_EXIT_OR_EVICT instead of setting checkpoint_exit_code. This will cause HTCondor to transfer your checkpoint file(s) (which you listed in transfer_output_files, as noted above) when the job is evicted. Of course, since this is the only time your checkpoint file(s) will be transferred, if the transfer fails, your job has to start over from the beginning. One reason file transfer on eviction fails is if it takes too long, so this method may not work if your transfer_output_files contain too much data.

Furthermore, eviction can happen at any time, including while the code is updating its checkpoint file(s). If the code does not update its checkpoint file(s) atomically, HTCondor will transfer the partially-updated checkpoint file(s), potentially overwriting the previous, complete one(s); this will probably prevent the code from picking up where it left off.

In some cases, you can work around this problem by using a wrapper script. The idea is that renaming a file is an atomic operation, so if your code writes checkpoints to one file, call it checkpoint, your wrapper script – when it detects that the checkpoint is complete – would rename that file checkpoint.atomic. That way, checkpoint.atomic always has a complete checkpoint in it. With a such a script, instead of putting checkpoint in transfer_output_files, you would put checkpoint.atomic, and HTCondor would never see a partially-complete checkpoint file. (The script would also, of course, have to copy checkpoint.atomic to checkpoint before running the code.)

Manual Transfers

If you’re comfortable with programming, instead of running a job with checkpoint_exit_code, you could use condor_chirp, or other tools, to manage your checkpoint file(s). Your executable would be responsible for downloading the checkpoint file(s) on start-up, and periodically uploading the checkpoint file(s) during execution. We don’t recommend you do this for the same reasons we recommend against managing your own input and output file transfers.

Early Checkpoint Exits

If your executable’s natural checkpoint interval is half or more of your pool’s max job runtime, it may make sense to checkpoint and then immediately ask to be rescheduled, rather than lower your user priority doing work you know will be thrown away. In this case, you can use the OnExitRemove job attribute to determine if your job should be rescheduled after exiting. Don’t set ON_EXIT_OR_EVICT, and don’t set +WantFTOnCheckpoint; just have the job exit with a unique code after its checkpoint.

Signals

Signals offer additional options for running self-checkpointing jobs. If you’re not familiar with signals, this section may not make sense to you.

Periodic Signals

HTCondor supports transferring checkpoint file(s) for an executable which takes a checkpoint when sent a particular signal, if the executable then exits in a unique way. Set +WantCheckpointSignal to TRUE to periodically receive checkpoint signals, and +CheckpointSig to specify which one. (The interval is specified by the administrator of the execute machine.) The unique way may be a specific exit code, for which you would set checkpoint_exit_code, or a signal, for which you would set +SuccessCheckpointExitBySignal to TRUE and +SuccessCheckpointExitSignal to the particular signal. (If you do not set checkpoint_exit_code, you must set +WantFTOnCheckpoint.)

Delayed Transfer with Signals

This method is very similar to but riskier than delayed transfers, because in addition to delaying the transfer of the checkpoint files(s), it also delays their creation. Thus, this option should almost never be used; if taking and transferring your checkpoint file(s) is fast enough to reliably complete during an eviction, you’re not losing much by doing so periodically, and it’s unlikely that a code which takes small checkpoints quickly takes a long time to start up. However, this method will work even with very old version of HTCondor.

To use this method, set when_to_transfer_output to ON_EXIT_OR_EVICT and KillSig to the particular signal that causes your job to checkpoint.