You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 97 Next »

I list each task here and what is necessary to run them in HTCondor.  I am assuming this will be running without a shared filesystem and also without access to NRAO filesystems.  So any call to /lustre/aoc or /users/<username> or other such things need to be altered to be site agnostic.

Every DAG or task creates .log, .out and maybe .png files that we want to keep.  Also, .last files like tclean.last are often created and are not necessary but can be usefull for debugging.  I assume that almost all tasks require the Measurement Set (MS).  I question what tasks actually modify the MS.  The reference to datacolumn is stating which column it should read from,  it does not imply any change to the MS.  Task07 sets savemodel='datacolumn' which actually modifies the MS.

This document is not complete.  I am sure I am missing inputs and perhaps outputs as well.

In this document, "data" when referenced as an input or an output is a directory containing the Measurement Set (E.g. VLASS1.2.sb36491855.eb36574404.58585.53016267361_split.ms/).  The jobs are run in the working directory so any file references are relative to that.

How do we handle the want to start a job at a given task?  For example, say a job ran to completion but you want to re-run the job after altering something in task17.  It would be unfortunate to have to run tasks 1 through 16.  It would be better to start and task17 and run through to the end of task25.  To do this requires saving the output of each task.  But how?  Incremental or Differential?  Using prolog and epilog scripts? Other?

Does run_tclean() need just the .psf directories or does it need more than that?  Tclean will need all image types (suffixes) for the named image.  For instance Task01 makes a set of 'iter0' images,  task04 makes an 'iter1' set of images.  Task 5 references both.  It would be acceptable to pass images of iter0* and iter1* but in practice it only needs the PSF from both so something like iter0*.psf and iter1*.psf  should work.

Is it safe to assume I don't need to transfer lockfiles like table.lock even if they have been modified?


How do we transfer input files for each DAG?

  • explicity list every file/directory in transfer_input_files (it doens't grok regexps).  This would be a large list .  E.g.
    • transfer_input_files = "working/VIP_iter0.gridwt, working/VIP_iter0.pb.tt0, working/VIP_iter0.psf.tt0, working/VIP_iter0.psf.tt1, working/VIP_iter0.psf.tt2, working/VIP_iter0.sumwt.tt0, working/VIP_iter0.sumwt.tt1, working/VIP_iter0.sumwt.tt2, working/VIP_iter0.weight.tt0, working/VIP_iter0.weight.tt1, working/VIP_iter0.weight.tt2"
  • Can transfer_input_fies take a manifest?  E.g a file containing the list of files to transfer
  • Make a temporary director on the submit host, and transfer that (possibly tarring it up)
  • Set the inputs and outputs for both data and working as a variables in the unified DAG file.  The task.sh script uses rsync to merge the various data_inputs together into one data directory and the various working_inputs together into one working directory.  Then at the end, task.sh moves data to data-<dagstep> and working to working-<dagstep> and the appropriate dirs/files from these are transferred back to the submithost.  The result of all this is that the data needed as an input for a step (E.g. Task08) may need to be combined from multiple places (initial data and data output from Task07)

To Do

  • DONE: write .log, .out and .png files one level up so they are not in the working directory and therefore not copied to execute hosts.
  • DONE: add rm -f *.last to the sh script?
  • DONE: Task24 and Task25 swapped with 8cores so they need to run with fewer cores.  So I may need to make another variable to pass to the sh script for this.
  • DONE: re-create my DOT graph after finishing task01-25-parallel-dag4. Also make a PDF instead of a PS.
  • DONE: Craft parallel-dag5 with the concept of running tclean at CHTC and everything else locally.
  • DONE: I don't like using the name Task as that has meaning to CASA.  A better term might be Step as in DAG Step.
  • Task19 needs to be unwraveled from NRAO filesystems.  Actually maybe not.  This can be a task we always run here.  But it does need to move into a VLASS area instead of Josh's home acocunt.
  • Figure out how to not copy SYSPOWER in the MS.  Presumably we can just cp /dev/null SYSPOWER/table.f0 and cp /dev/null SYSPOWER/table.f0i
  • Setup testpost-serv-1 with Lustre access over IB so we can start submitting to CHTC.


Possible Improvements

  • Task04 could perhaps be run concurrently with Task01 as long as Task05 is run after both.
  • Task13 could perhaps be run concurrently with Task12.
  • Task14 could perhaps be run concurrently with Taks12 and Task13 as long as Task15 is run after both Task12 and Task14.
  • Task22 could perhaps be run concurrently with Task12 or later as long as Task23 is run after both Task22 and Task13.


Task01 - Step01

Doesn't alter the MS

run_tclean( 'iter0', cfcache=cfcache_nowb, robust=-2.0, uvtaper='3arcsec', calcres=False  )

Task02

This tasks creates VIP_iter0b.* but I don't see those files ever referenced in this script again.  What does this task do that is necessary to other tasks?  Josh Marvil said that this is a leftover task and can be removed.

Doesn't alter the MS

run_tclean( 'iter0b', cfcache=cfcache_nowb, calcres=False  )


Task03

This task doesn't parallelize and only takes tens of seconds to run.  Should this be stuck on the end of task01?

Doesn't alter the MS

mask_from_catalog(inext=inext,outext="QLcatmask.mask",catalog_search_size=1.5,catalog_fits_file='../VLASS1Q.fits')

  • input: ../data
  • input: ../VLASS1Q.fits, VIP_iter0.psf.tt0
  • output: mask_from_cat.crtf, VIP_QLcatmask.mask


Task04 - Step04

Doesn't alter the MS

This task could possibly run at the same time as Task01 except that I have combined this with Task05 which requires both Task01 and Task04.

run_tclean( 'iter1', robust=-2.0, uvtaper="3arcsec"  )

  • input: ../data
  • output: VIP_iter1.*


Task05

Doesn't alter the MS

replace_psf('iter1','iter0')

This is just some python that deletes VIP_iter1.psf.* and copies VIP_iter0.psf.* to VIP_iter1.psf.*.  It is inefficient to ever make this task be its own DAG.  I suggest it always be in the same DAG as Task04.  Will produce an error because *.workdirectory doesn't exist but that error is ignorable.

  • input: VIP_iter0.psf.*, VIP_iter1.psf.*
  • output: VIP_iter1.psf.*


Task06 - Step06

Doesn't alter the MS

run_tclean( 'iter1', robust=-2.0, uvtaper="3arcsec", niter=20000, nsigma=5.0, mask="QLcatmask.mask", calcres=False, calcpsf=False  )

  • input: ../data
  • input: VIP_iter1.*, VIP_QLcatmask.mask
  • output: VIP_iter1.*


Task07 - Step07

Alters the MS

run_tclean( 'iter1', calcres=False, calcpsf=False, savemodel='modelcolumn', parallel=False  )

Task08 - Step08

Alters the MS

Tasks 08, 09, 10 and 11 take only minutes to run so could be combined into one DAG step.

flagdata(vis=vis, mode='rflag', datacolumn='residual_data',timedev='tdev.txt',freqdev='fdev.txt',action='calculate')

replace_rflag_levels()

flagdata(vis=vis, mode='rflag', datacolumn='residual_data',timedev='tdev.txt',freqdev='fdev.txt',action='apply',extendflags=False)

flagdata(vis=vis, mode='extend', extendpols=True, growaround=True)


Task09

Alters the MS

statwt(vis=vis,combine='field,scan,state,corr',chanbin=1,timebin='1yr', datacolumn='residual_data' )


Task10

Doesn't alter the MS

gaincal(vis=vis,caltable='g.0',gaintype='T',calmode='p',refant='0',combine='field,spw',minsnr=5)

  • input: ../data
  • output: g.0


Task11

Alters the MS

applycal(vis=vis,calwt=False,applymode='calonly',gaintable='g.0',spwmap=18*[2], interp='nearest')


Task12 - Step12

Doesn't alter the MS

run_tclean( 'iter0c', datacolumn='corrected', cfcache=cfcache_nowb, robust=-2.0, uvtaper='3arcsec', calcres=False  )

  • input: ../data
  • output: VIP_iter0c.*


Task13 - Step13

Doesn't alter the MS

Could this run in parallel with one or more previous run_tclean calls like Task12?

run_tclean( 'iter0d', datacolumn='corrected', cfcache=cfcache_nowb, calcres=False  )

  • input: ../data
  • output: VIP_iter0d.*


Task14 - Step14

Doesn't alter the MS

This task could possibly run at the same time as Task12 or Task13 except that I have combined this with Task15 which requires both Task14 and Task12.

run_tclean( 'iter1b', datacolumn='corrected', robust=-2.0, uvtaper="3arcsec" )

  • input: ../data
  • output: VIP_iter1b.*


Task15

Doesn't alter the MS

replace_psf('iter1b','iter0c')

This is just some python that deletes VIP_iter1b.psf.* and copies VIP_iter0c.psf.* to VIP_iter1b.psf.*.  It is inefficient to ever make this task be its own DAG.  I suggest it always be in the same DAG as Task14.  Will produce an error because *.workdirectory doesn't exist but that error is ignorable.

  • input: VIP_iter1b.psf.*, VIP_iter0c.psf.*
  • output: VIP_iter1b.psf.*


Task16 - Step16

Doesn't alter the MS

run_tclean( 'iter1b', datacolumn='corrected', robust=-2.0, uvtaper="3arcsec", niter=20000, nsigma=5.0, mask="QLcatmask.mask", calcres=False, calcpsf=False  )

  • input: ../data
  • input: VIP_iter1b.*, VIP_QLcatmask.mask
  • output: VIP_iter1b.*


Task17 - Step17

Doesn't alter the MS

imsmooth(imagename=imagename_base+"iter1b.image.tt0", major='5arcsec', minor='5arcsec', pa='0deg', outfile=imagename_base+"iter1b.image.smooth5.tt0")

  • input: ../data
  • input: VIP_iter1b.image.tt0
  • output: VIP_iter1b.image.smooth5.tt0


Task18

Doesn't alter the MS

exportfits(imagename=imagename_base+"iter1b.image.smooth5.tt0", fitsimage=imagename_base+"iter1b.image.smooth5.fits")

  • input: ../data
  • input: VIP_iter1b.image.smooth5.tt0
  • output: VIP_iter1b.image.smooth5.fits


Task19

This needs some modification. It calls a script from Josh's homedir and runs bdsf out of /lustre. Also, I have been unable to run this task by itself.  I get the following errors.  I am going to combine this with tasks17, 18, 20 and 21 so it isn't an issue right now.

2020-07-29 16:13:07     SEVERE  exportfits::image::tofits (file ../../tools/images/image_cmpt.cc, line 6211)    Exception Reported: Exception: File VIP_iter1b.image.smooth5.fits exists, and the user does not want to remove it..
2020-07-29 16:13:07     SEVERE  exportfits::image::tofits (file ../../tools/images/image_cmpt.cc, line 6211)+   ... thrown by static void casa::ImageFactory::_checkOutfile(const casacore::String&, casacore::Bool) at File: ../../imageanalysis/ImageAnalysis/ImageFactory2.cc, line: 568
2020-07-29 16:13:07     SEVERE  exportfits::::@testpost001:MPIClient    An error occurred running task exportfits.

I de-linked everything in bdsf_1.8.15-py27-rh7-env to make it more portable but then would sometimes get Failed to transfer files errors in the condor.log which seems to be related to the number of files which was about 12,000. I could reproduce the problem by transferring a directory with about 10,008 files in it but not repeatedly reproduce it.  So there is something else going wrong.  Anyway, since these tasks are not planned to run at CHTC I am going to go back to calling this out of Josh's home account.  A longer-term solution might be to move it out of josh's area and into VLASS or tar up the portable directory and copy that instead.


Doesn't alter the MS

subprocess.call(['/users/jmarvil/scripts/run_bdsf.py', imagename_base+'iter1b.image.smooth5.fits'],env={'PYTHONPATH':''})

  • input: ../data
  • input: VIP_iter1b.image.smooth5.fits
  • input: ??
  • output:
    • VIP_iter1b.image.smooth5.cat.ds9.reg
    • VIP_iter1b.image.smooth5.cat.fits
    • VIP_iter1b.image.smooth5.fits.island.mask
    • VIP_iter1b.image.smooth5.fits.pybdsf.log
    • VIP_iter1b.image.smooth5.fits.rms
    • VIP_iter1b.image.smooth5.fits


Task20

This needs iter0.psf.tt0 which is set to the variable inext.  VIP_iter0 was copied to VIP_iter1 back in Task05

Doesn't alter the MS

edit_pybdsf_islands(catalog_fits_file=imagename_base+'iter1b.image.smooth5.cat.fits')

mask_from_catalog(inext=inext,outext="secondmask.mask",catalog_fits_file=imagename_base+'iter1b.image.smooth5.cat.edited.fits', catalog_search_size=1.5)

  • input: VIP_iter1b.image.smooth5.cat.fits
  • input: VIP_iter1b.image.smooth5.cat.edited.fits, VIP_iter0.psf.tt0
  • output: secondmask.mask


Task21

Doesn't alter the MS

immath(imagename=[imagename_base+'secondmask.mask',imagename_base+'QLcatmask.mask'],expr='IM0+IM1',outfile=imagename_base+'sum_of_masks.mask')

im.mask(image=imagename_base+'sum_of_masks.mask',mask=imagename_base+'combined.mask',threshold=0.5)

  • input: secondmask.mask, VIP_QLcatmask.mask
  • output: sum_of_masks.mask
  • input: sum_of_masks.mask
  • output: VIP_combined.mask


Task22 - Step22

Doesn't alter the MS

As far as I can tell at this point, ../data has not changed since Task11 (applycal).

Could this run in parallel with one or more previous run_tclean calls like Task16?

run_tclean( 'iter2', datacolumn='corrected' )

  • input: ../data
  • output: VIP_iter2.*


Task23

Doens't alter the MS

replace_psf('iter2', 'iter0d')

This is just some python that deletes VIP_iter2.psf.* and copies VIP_iter0d.psf.* to VIP_iter2.psf.*.  It is inefficient to ever make this task be its own DAG.  I suggest it always be in the same DAG as Task22.

  • input: VIP_iter2.psf.*, VIP_iter0d.psf.*
  • output: VIP_iter2.psf.*


Task24

Doesn't alter the MS

At this point I think we are using iter2's image with iter0d's psf.

I ran this on a node with 512GB, asking for 500GB, 8 cores and using mpicasa -n 9 this task swapped.  So I am restarting it with 4 cores and -n 5

run_tclean( 'iter2', datacolumn='corrected', scales=[0,5,12], nsigma=3.0, niter=20000, cycleniter=3000, mask="QLcatmask.mask", calcres=False, calcpsf=False  )

  • input: ../data
  • input: VIP_iter2.*, VIP_QLcatmask.mask
  • output: VIP_iter2.*


Task25

os.system('rm -rf *.workdirectory')

os.mkdir('iter2_intermediate_results')

os.system('cp -r *iter2* iter2_intermediate_results')

shutil.rmtree(imagename_base+'iter2.mask')

shutil.copytree(imagename_base+'combined.mask',imagename_base+'iter2.mask')

run_tclean( 'iter2', datacolumn='corrected', scales=[0,5,12], nsigma=3.0, niter=20000, cycleniter=3000, mask="", calcres=False, calcpsf=False  )

This does some file cleaning and then runs run_tclean.  Where do we want to do that file cleaning?  In the previous task?  On the submit host?

  • input: ../data
  • input: VIP_iter2.*, VIP_combined.mask
  • output: VIP_iter2.*

















  • No labels