checkpoint_techniques_on_compute_canada_clusters

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision		 Previous revision
checkpoint_techniques_on_compute_canada_clusters [2015/03/27 15:05] – [Manual checkpoints] 132.216.122.26checkpoint_techniques_on_compute_canada_clusters [2024/03/26 13:52] (current) – external edit 127.0.0.1
Line 1: Line 1:
-These are the notes for the Checkpoint Techniques workshop I attended on March 26th, 2015. Might be useful for people who want to learn how to code this on their own programs. Please don't hesitate to edit this page if you feel I left something out, you want to add something on your own or my English sounds too Spanish.+These are the notes for the Checkpoint Techniques workshop I attended on March 26th, 2015 (the workshop materials can be found [[http://www.hpc.mcgill.ca/index.php/training#chkpt|here]].) Might be useful for people who want to learn how to code this on their own programs. Please don't hesitate to edit this page if you feel I left something out, you want to add something on your own or my English sounds funny.
  
 ===== Random stuff ===== ===== Random stuff =====
  
-* Maximum ''wall'' time on ''guillimin'' is 30 days. You can reach this limit without the need to ask any administrator (someone told me the limit was 48 hours before, but now it does not seem to be the case.) Also, it cannot be raised, as the scheduler will kill jobs longer than that. If you want to run something that will take longer, checkpointing comes in handy.+  * Maximum ''wall'' time on ''guillimin'' is 30 days. You can reach this limit without the need to ask any administrator (someone told me the limit used to be 48 hours, but now it does not seem to be the case.) Also, it cannot be raised higher, as the scheduler will kill jobs longer than that. If you want to run something that will take longer, checkpointing comes in handy.
  
-* There is a [[https://wiki.calculquebec.ca/w/Accueil|Calcul Québec wiki]].+  * There is a [[https://wiki.calculquebec.ca/w/Accueil|Calcul Québec wiki]].
  
 ===== The problem ===== ===== The problem =====
Line 17: Line 17:
 ==== Manual checkpoints ==== ==== Manual checkpoints ====
  
-Manual checkpoints require you to modify your code to manually save your system state every N iterations (for example) and then, upon start, check if there is some previously saved state that can be used as the initial seed. In some cases this might not be possible, depending on your particular job (for instance, it might be a very long optimization process that depends on some internal hysteresis data that is not offered as output). The particular file format used for this task is up to you. In my case, I save intermediate parameter values in ''.Rdata'' format, as I am using R. The same goes for Matlab. If you are using Python, you might want to check [[https://docs.python.org/2/library/pickle.html|Pickle]], if you didn't know it already.+Manual checkpoints require you to modify your code to manually save your system state every N iterations (for example) and then, upon start, check if there is some previously saved state that can be used as the initial seed. In some cases this might not be possible, depending on your particular job (for instance, it might be a very long optimization process that depends on some internal hysteresis data that is not offered as output). The particular file format used for this task is up to you. In my case, I save intermediate parameter values in ''.Rdata'' format, as I am using R. The same goes for Matlab (''.mat''.) If you are using Python, you might want to check [[https://docs.python.org/2/library/pickle.html|Pickle]], if you didn't know it already.
  
 If you are using parallel programs, you might be interested in HDF5, which allows to read/write data in parallel environments. Check [[http://www.hdfgroup.org/HDF5|this webpage]] for more information. These libraries and programs are offered on ''guillimin'' simply by running ''module add HDF5/1.8.7-gcc'' (or ''module add HDF5/1.8.7-intel'', depending on the compiler you are using.) If you are using parallel programs, you might be interested in HDF5, which allows to read/write data in parallel environments. Check [[http://www.hdfgroup.org/HDF5|this webpage]] for more information. These libraries and programs are offered on ''guillimin'' simply by running ''module add HDF5/1.8.7-gcc'' (or ''module add HDF5/1.8.7-intel'', depending on the compiler you are using.)
Line 45: Line 45:
 # 7779 by default, but if there are several DMTCP schedulers running on  # 7779 by default, but if there are several DMTCP schedulers running on 
 # the same node we will have problems. The best solution is to assign the # the same node we will have problems. The best solution is to assign the
-# port number manually.+# port number manually. Also, if PORT=0, a random unused port will be  
 +# chosen, which is probably better.
 PORT=7745 PORT=7745
  
Line 59: Line 60:
 else else
     # The -i switch tells dmtcp_launch the time in seconds between     # The -i switch tells dmtcp_launch the time in seconds between
-    # checkpoints.+    # checkpoints. Probably 60 is too small, so set it up accordingly.
     dmtcp_launch -i 60 -p ${PORT} ./my_job     dmtcp_launch -i 60 -p ${PORT} ./my_job
 fi fi
 </file> </file>
  
-That's it. DMTCP will regularly create system dumps and link ''dmtcp_restart_script.sh'' (created on the directory from which you run your job) to the relevant copy.+That's it; we will use this script to launch our job and to re-launch it if we need to. DMTCP will regularly create system dumps and link ''dmtcp_restart_script.sh'' (created on the directory from which you run your job) to the relevant copy.
  
 === What if I want to launch N jobs using DMTCP? Is that possible? === === What if I want to launch N jobs using DMTCP? Is that possible? ===
  
 Yes. You just have to remember to assign each job a different port number and run every job from its own subdirectory to avoid any conflict. Yes. You just have to remember to assign each job a different port number and run every job from its own subdirectory to avoid any conflict.
 +
 +Consider my particular case: I need to run several R programs (hundreds). Each script will run on one node. This is the Python code that I have used to generate the bash scripts that I will send to ''qsub'':
 +
 +
 +<file python generatordmtcp.py>
 +#!/usr/bin/env python
 +import os
 +
 +# New version of this script. Now we use DMTCP to launch
 +# the scripts.
 +
 +def chunks(l, n):
 +    """ Yield successive n-sized chunks from l.
 +    """
 +    for i in xrange(0, len(l), n):
 +        yield l[i:i+n]
 +
 +## MAIN ##
 +if (__name__ == "__main__"):
 +
 +    NPROCS = 12 
 +
 +    # Get list of scripts to run. They are files with both
 +    # 'modelimp1_global' and 'CLUSTER' on their names. Each
 +    # script is one job to launch.
 +    scripts = os.listdir('.')
 +    scripts = filter(lambda x: x.find('CLUSTER') != -1, scripts)
 +    scripts = filter(lambda x: x.find('modelimp1_global') != -1, scripts)
 +    scripts.sort()
 +    id = 0
 +    # We'll save temporary results in the projects directory, so we 
 +    # don't have to worry about quotas on the scratch one. Might need
 +    # these data for several weeks.
 +    optdir = "/gs/project/eim-670-aa/jmateos/esmglobalfit"
 +    # Port list for DMTCP
 +    ports = range(7701, 7713)
 +
 +    ## MAIN LOOP ##
 +    for batch in chunks(scripts, NPROCS):
 +        id = id + 1
 +        jobname = "esmglobal_%02d" % id
 +        btemp = """#!/bin/bash
 +#PBS -A eim-670-aa
 +#PBS -l nodes=1:ppn=%d
 +#PBS -l walltime=00:30:00
 +#PBS -V
 +#PBS -N %s
 +#PBS -o %s 
 +#PBS -e %s
 +
 +function rundmtcpjob () {
 +    jobfile=$1
 +    port=$2
 +
 +    jobname=$(basename ${jobfile} .R)
 +    optdir=/gs/project/eim-670-aa/jmateos/esmglobalfit
 +
 +    # Create job directory within ${optdir} and copy all files there
 +    # If it already exists, it might mean the script already run once,
 +    # so don't do anything.
 +    scdir=${optdir}/${jobname}
 +    if [ ! -e ${scdir} ]
 +    then
 +        mkdir ${scdir}
 +        cp -va * ${scdir}
 +    fi
 +    
 +    # Move to $scdir and run the script using dmtcp_launch, as in the
 +    # workshop. Will use that directory as the temporary one.
 +    cd ${scdir}
 +    if [ -e "dmtcp_restart_script.sh" ]
 +    then
 +        ./dmtcp_restart_script.sh -p ${port} -h $(hostname)
 +    else
 +        dmtcp_launch -i 86400 -p ${port} R CMD BATCH ${jobfile}
 +    fi
 +}
 +
 +cd /home/jmateos/code/devmodel/devmodelR
 +
 +""" % (len(batch), jobname, \
 +       optdir + '/' + jobname + '.log', \
 +       optdir + '/' + jobname + '.err')
 +
 +        jobsfile = jobname + '.sh'
 +        f = open(jobsfile, 'w')
 +        f.write(btemp)
 +        for i in range(len(batch)):
 +            line = "rundmtcpjob %s %d &\n" % (batch[i], ports[i])
 +            f.write(line)
 +        f.write("wait\n")
 +        f.close()
 +        os.chmod(jobsfile, 0755)
 +    # end for loop
 +</file>
 +
 +In the end, this script generates a bunch of ''.sh'' files (as many as needed), splits the input files in chunks of ''NPROCS'' within the same node and makes the scripts executable. The only thing needed afterwards is to check that all the parameters are correct and ''qsub'' them. It will use the personal project space (hardcoded, you will need to change this) to create task subdirectories for DMTCP.
 +
 +**Currently this is not working as expected; for some unknown reason, only 2 random jobs get re-started. I have contacted Calcul Québec about this and they should reply shortly. I will update this page with a bug-free script (or whatever solution they give me.)** 
 +
 +**Update 2: they did not reply.**
  • checkpoint_techniques_on_compute_canada_clusters.1427468740.txt.gz
  • Last modified: 2024/03/26 13:52
  • (external edit)