Hello all,

    You can get this documentation and relevant data sets from

ftp ftp.cv.nrao.edu
cd /pub/NRAO-staff/efomalon/AIPS++

    or

ftp://ftp.cv.nrao.edu/NRAO-staff/efomalon/aips++/

    From this wide-field imaging/deconvolution test, I find that AIPS
runs 3 times faster than AIPS++.  Care was taken to make sure that the
convolution process was as similar as possible in both systems.  The
two major parts of the processing--(1) making the residual beam and
image; (2) finding and subtracting out the clean components--were both
consistent with this 3:1 ratio.  The other parts of the processing
were insignificant compared to the above two step.

    The two images agree remarkably well, as the difference image
shows.  The very subtle difference may be caused by the somewhat
different weighting algorithms.

    Please play around with the data set.  It is representative of a
realistic data set associated with sensitive, wide-field imaging that
is needed for the VLA, EVLA and ALMA.  If there are ways of speeding
up the processing, please let me know.


Files:

README:  This document

RR2: 3 million data points, each with 7-channels a pseudo-continuum
data set from the VLA A and B-configurations.  The data were
calibrated, edited and self-calibrated about two years ago in AIPS.
The bright sources (up to 20 mJy) were subtracted out early in the
calibration process.  It is a full-primary beam deep image at 1.4 GHz
with an rms noise of 10 uJy, made from 100 hours of VLA time.  This
data set contains only 1 IF and stokes RR.  The 2 IF's and 2 stokes
were processed separately, but this data base contains only one of the
four.

RR2_AIPS: The wide-field clean image made in AIPS.

RR2_AIPSPP:  The wide-field clean image made in AIPS++.

RR2_DIFF:  The difference between the two images.

zclean_script.g:  The aips++ and aips scripts used for the imaging.
         
logfile:  The aips++ and aips log files

-------------------------------------------------------------------

Imaging/deconvolution parameters:

1.  Made 8192x8192 image with 0.5" pixels which covers entire primary beam
    5x5 = 25 facets used in aips++; 31 facets used in AIPS (facets are 
defined
    with different overlap properties).

2.  Peak on image is 900 uJy, rms noise is 9.9 uJy, so
    dynamic range is <100:1 - not all that high.

3.  Ran wide-field Clark cleaing in aips and aips++ down to a residual 
of 50 uJy.
    Used ROBUST = 1 with 1.8"x1.8" clean beam
    Two major cycles were used in both cases, and the number of clean 
components
    found were approximately the same.


BENCH MARK OF WIDE FIELD IMAGING for major reduction step:

AIPS++ deep imaging

   35m to make 25 beams
   36m to make 25 residual image
   55m to find 415 components in 12 fields and gridded subtract to reach 
0.0002 Jy level
               (4.5m per facet)
   41m to make 25 new residual images
  125m to find ~4700 components to in 25 fields and gridded sub to reach 
0.00005 Jy level
               (5m per facet)
   41m to make new residual image
    2m to do the restoration
------
  335m total

AIPS deep imaging

   10m to make 31 beams
   12m to make 31 residual image
   20m to find 424 components in 11 fields and gridded subtract to reach 
0.0002 Jy level
               (1.8m per facet)
   12m to make 31 residuals
   35m to find 4263 components in 27 fields and gridded subtract to 
reach 0.00005 Jy level
               (1.3m per facet to process)
   12m to make 31 new residual images
    1m to make new residual image
    6m to glues all facets together (flatn)
------
  108m total         

The timing for the basic steps are:

        STEP             AIPS++ time            AIPS  time

      Make beam facet    1.4 m/facet            0.4 m/facet
      Make image facet   1.4 m/facet            0.4 m/facet
      Find components
         to subtract     4.5 m/facet            1.5 m/facet

------------------------------------------------------------------------------

                           CONCLUSION

AIPS is about 3 times faster than AIPS++ for this widefield example.
The computer has 4 Gigs of memory.  I set the system.resources.memory: 900.
However, on execution 400 Megs at most was the largest amount of memory 
used.


------------------------------------------------------------------------------