Документ взят из кэша поисковой машины. Адрес
оригинального документа
: http://www.adass.org/adass/proceedings/adass98/youngw/
Дата изменения: Sat Jul 17 00:46:54 1999 Дата индексирования: Tue Oct 2 06:51:55 2012 Кодировка: Поисковые слова: http news.cosmoport.com 2001 09 19 2.htm |
Data reduction in radio astronomy has several processing steps which are amenable to parallelization. Spectral-line deconvolution is an obvious candidate since processing can be carried out on each plane independently.
We demonstrate here a ``parallelized'' Clark CLEAN (Clark 1980) algorithm implemented in the AIPS++ environment.
The demonstration by necessity uses small image cubes (512 x 512 pixels by 100 channels) for timeliness. We make significant time-savings gains when processing large image cubes with many planes and pixels.
We chose an Algorithm-Applicator scheme for the embarrassingly parallel problems. To communicate between the Algorithm and Applicators we use a PTransport object. This scheme is well suited to the AIPS++ programming environment.
The Algorithm runs n-copies of itself, each are independent of its siblings (in our case, the deconvolution of one plane in a dirty image). The Applicator is the controller, setting up the problem, sends data to the Algorithm processes, and receives the results.
PTransport is the conduit for transferring data between the Algorithms and the Applicator. Using this design provides a way to take advantage of parallel processing if it's available. Object diagrams of the Applicator, Algorithm, and PTransport classes are shown in Fig. 1.
![]() |
We use a simple master/slave scheme for doing the parallelization. The Applicator (pimager) does the bookkeeping and controls the IO to and from the AIPS++ system. The Algorithm (Clark CLEAN) processes receive a plane from the dirty image and deconvolution parameters from the Applicator, deconvolves the image, and send the results back to the Applicator.
The Applicator determines what type of PTtransport to use at run time, parallel, serial or in the future threads. Currently we have implemented serial and MPI PTransports1. A thread based PTransport will be implemented (once the AIPS++ libraries are thread-safe).
Our system is evolving. To process on the ``large'' multiprocessor SGI systems at NCSA we create a command script and pass it into a batch queue. Notification of when the processing is complete is done via e-mail.
Preliminary testing of parallelization in AIPS++ show promising speed ups for multi-channel deconvolution. Figure 4 shows the speed up falling away from linear at 15 processors. We are investigating why this is happening.
![]() |
Dave Westpfahl (New Mexico Tech), supplied us with four pointings of VLA HI data of M33 (B, C, and D arrays) to create the images. Data in the demo are from one of these pointings.
We intend to combine the four pointings into a mosaic of 6000 x 6000 pixels (about two degrees on a side) by 140 channels.
Clark, B. G. 1980, A&A89, 377