The batch processor allows Automatic Run Processor (or ARP, for short, because I want that to catch on) is a web service that allows for automatic workflows and for the easier submission of batch jobs via a web interface. A script that submits the batch job (to allow for more customization in this command) is all that is needed for this system to work.
Table of Contents |
---|
To use this system, choose an experiment from https://pswww.slac.stanford.edu/lgbk/lgbk/experiments.
Under the Workflow dropdown, select Definitions. The Workflow Definitions tab is the location where scripts are defined; scripts are registered with a unique name. Any number of scripts can be registered. To register a new script, please click on the + button on the top right hand corner.
A hashtag is created which represents some script via the script's absolute path. This script contains the bsub command so that extra tags can be added and more customization of the bsub command is available. This hashtag can then be applied to events as needed.
Below is a simple example of a possible bash script. This script runs test.py
and passes it any arguments that is passed to itself through "$@"
and the log files are save in the logs
directory with their job ID as the log file name. The python code can do analysis on the run this script was applied to
...
A unique name given to a registered script; the same script can be registered under different names with different parameters.
The absolute path to the batch script. This script must contain the batch job submission command (sbatch for SLURM). It gives the user the ability to customize the the batch submission. Overall, it can act as a wrapper for the code that will do the analysis on the data along with submitting the job.
The parameters that will be passed to the executable as command line arguments. These parameters can be used as parameters to the sbatch command to specify the queue, number of cores etc. Or; they can be used to customize the script execution. In addition, details of the batch job are made available as as environment variables.
ARP_UPDATE_COUNTERS - This is a URL that can be used to update the progress of the job. These updates are also automatically reflected in the UI. In previous releases, this was called JID_UPDATE_COUNTERS.
This defines where the analysis is done. While many experiments prefer to use the SLAC psana cluster (SLAC) or the SRCF (SRCF_FFB) to perform their analysis, others prefer to use HPC facilities like NERSC to perform their analysis.
This defines the event that in the data management system that kicks off the job submission.
If the job is automatically triggered, it will be executed as this user. If the job is manually triggered; it will be executed as the user triggering the job manually. This is set when creating the job definition and cannot be changed.
Under the Workflow dropdown, select Control to create and check the status of your analysis jobs. The Control tab is where job definitions defined in the Definitions tab may be applied to experiment runs. An entry is automatically created for jobs that are triggered automatically. To manually trigger a job, in the drop-down menu of the Job column, select the job. A job can be triggered any number of times; each execution has a separate row in the UI.
These are the different statuses that a job can have -
There are four different actions which can be applied to a script. They do the following if pressed:
- Attempt to kill the job. A green success statement will appear near the top-right of the page if the job is killed successfully and a red failure statement will appear if the job is not killed successfully.
- Returns the log file for the job. If there is no log file or if no log file could be found, it will return blank.
- Returns details for the current job by invoking the the appropriate job details command in the HPC workload management infrastructure.
- Delete the job execution from the run. Note: this does not kill the job, it only removes it from the webpage.
This is a customizable column which can be used by the script executable to report progress. The script executable reports progress by posting JSON to a URL that is available as the environment variable JID_UPDATE_COUNTERS.
For example, to update the status of the job using bash, one can use
Code Block | ||
---|---|---|
| ||
curl -s -XPOST ${JID_UPDATE_COUNTERS} -H "Content-Type: application/json" -d '[ {"key": "<b>LoopCount</b>", "value": "'"${i}"'" } ]' |
In Python, one can use
Code Block | ||
---|---|---|
| ||
import os
import requests
requests.post(os.environ["JID_UPDATE_COUNTERS"], json=[ {"key": "<b>LoopCount</b>", "value": "75" } ]) |
The executable script used in the workflow definition should be used primarily to set up the environment etc and submit the analysis script to the HPC workload management infrastructure. For example, a simple executable script that uses SLURM's sbatch to submit the analysis script is available here - /reg/g/psdm/tutorials/batchprocessing/arp_submit.sh
Code Block |
---|
#!/bin/bash
source /reg/g/psdm/etc/psconda.sh
ABS_PATH=/reg/g/psdm/tutorials/batchprocessing
sbatch --nodes=2 --partition=psanaq --time=5 --output="arp_example_${RUN_NUM}_%j.log" $ABS_PATH/arp_actual.py "$@" |
This script will submit /reg/g/psdm/tutorials/batchprocessing/arp_actual.py. /reg/g/psdm/tutorials/batchprocessing/arp_actual.py will be passed the parameters as command line arguments and will inherit the EXPERIMENT, RUN_NUM and JID_UPDATE_COUNTERS environment variables.
Log files:
→ If the --output parameter is not specified to sbatch, then SLURM will store the log output in /reg/d/psdm/dia/diadaq13/scratch/<slurm_job_id>.out
→ In the example above, the log output will be sent to to the default working folder for the job; which is the scratch folder but the file name will be generated using the run number and the job id. For example, the log file for run 25 job id 409327. will be send to /reg/d/psdm/dia/diadaq13/scratch/arp_example_25_409327.log
→ To avoid cluttering the scratch folder, one can use an absolute path in the --output command to specify an alternate location for the job log files. See the "filename pattern" in the sbatch man page for more details.
This Python script is the code that will do analysis and whatever is necessary on the run data. Since this is just an example, the Python script, arp_actual.py, doesn't get that involved. It is shown below.
Code Block |
---|
#!/usr/bin/env python
import os
import sys
import requests
import time
import datetime
import logging
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger(__name__)
logger.debug("In the arp_actual script - current time is %s", datetime.datetime.now().strftime("%c"))
for k, v in sorted(os.environ.items()):
logger.debug("%s=%s", k, v)
## Fetch the URL to post progress updates
update_url = os.environ.get('JID_UPDATE_COUNTERS')
logger.debug("The URL to post updates is %s", update_url)
# These are the parameters that are passed in
logger.debug("The parameters passed into the script are %s", " ".join(sys.argv))
loop_count = 20
try:
loop_count = int(sys.argv[1])
except:
pass
## Run a loop, sleep a second, then POST
for i in range(loop_count):
time.sleep(1)
logger.debug("Posting for step %s", i)
requests.post(update_url, json=[{"key": "<strong>Counter</strong>", "value" : "<span style='color: red'>{0}</span>".format(i+1)}, {"key": "<strong>Current time</strong>", "value": "<span style='color: blue'>{0}</span>".format(datetime.datetime.now().strftime("%c"))}])
logger.debug("Done with job execution")
|
The logger.debug statements are sent to the job's log file. Note, one can form sbatch commands where the log output is not sent to a logfile and is instead sent as an email. Part of an example log file output is shown below.
No Format |
---|
DEBUG:__main__:In the arp_actual script - current time is Thu Apr 16 11:12:40 2020
...
DEBUG:__main__:EXPERIMENT=diadaq13
...
DEBUG:__main__:JID_UPDATE_COUNTERS=https://pswww.slac.stanford.edu/ws/jid_slac/jid/ws/replace_counters/5e98a01143a11e512cb7c8ca
...
DEBUG:__main__:RUN_NUM=26
...
DEBUG:__main__:The parameters passed into the script are /reg/g/psdm/tutorials/batchprocessing/arp_actual.py 100
DEBUG:__main__:Posting for step 0
DEBUG:urllib3.connectionpool:Starting new HTTPS connection (1): pswww.slac.stanford.edu:443
DEBUG:urllib3.connectionpool:https://pswww.slac.stanford.edu:443 "POST /ws/jid_slac/jid/ws/replace_counters/5e98a01143a11e512cb7c8ca HTTP/1.1" 200 195
DEBUG:__main__:Posting for step 1
...
DEBUG:__main__:Done with job execution
... |
Is it possible to submit more than one job per run?
Can a submitted job submit other subjobs?
When using the 'kill' option, how does ARP know which jobs to kill?
As far as I understand there is a json entry for each line which stores info, can one access this json entry somehow?