From efomalon@cv3.cv.nrao.edu Thu Aug 16 12:29:19 2001 Date: Thu, 16 Aug 2001 14:22:56 -0400 From: Ed Fomalont To: Athol Kemball Cc: fowen@zia.aoc.NRAO.EDU, aips2-naug@zia.aoc.NRAO.EDU Subject: Re: AIPS - AIPS++ Imaging test Hi Athol, et al First, I will be arriving in Socorro on Saturday and I hope I can work with you and Kumar on Monday concerning the timing of the mapping and deconvolution. Since most of the other aips++ testers are working on u-v data calibration, I decided to go ahead and try some of the more sophisticated imaging and deconvolution software on moderately large data bases. Aips++ has some imaging and deconvolution features which are not in aips, or are awkward to use in aips. One of the more important is deconvolution using the known beam-squint in the VLA primary beam. However, I am first starting with the most basic imaging and deconvolution packages and trying to work up to more complexity. I am now using a VLA 1.4 GHz data base with lots of sources over a wide field, so that the imaging quality and efficiency can be measured at a somewhat deeper level. The data were calibrated and editted in AIPS. One phase selfcalibration was also done in AIPS, so the data quality should be very good. I reran a few aips vs aips++ comparisons on my linux machine with 392 Megs of memory. I set in my .aipsrc file. I tried 392, but the imaging crashed with the usual memory kind of fault. All files are resident on my machine. The timing tests are (in a relatively empty machine): cpu run 2048x2048 image, 2000 clean iterations, no boxes time time briggs weighting, robust=1 aips using IMAGR 220s 380s aips++ clark clean with restore 632s 800s aips++ wfclark clean (256M memory) 1360s 1950s aips++ wfclark clean (default memory=64k?) 1660s 2300s Hence, adding a larger memory allocation in the .aipsrc file did help a bit. The images are about the same quality. This field really should be done with many facets (about 8 along each coordinate), but that is the next stage of testing. All of the images are about the same quality. Now, I realize that there may be lots of things one can twiddle to obtain better efficiency. As Debra mentioned, the values of these parameters should be automatically set when aips++ is executed, or well-documental so the user can specify the appropriate parameters. Whether the relatively poor aips++ efficiency I am getting is associated with a lack of fine-tuning the aips++ environment, or is more fundamental to the algorithms in the code is unclear. So, see you on Monday. Cheers, Ed