[daip] memory
Eric Greisen
egreisen at nrao.edu
Mon Jan 7 16:39:49 EST 2013
Marina Alexandrovna Kounkel wrote:
> Hello!
>
> I was curious, is there any way to feed aips more memory than 6114 MB at which setmaxap maxes out? Considering we have several dozens of GB available on a dedicated machine, this seems to be a rather small number. And/or, is there any other way at all to increase the speed of the operations?
>
> On a somewhat unrelated subject, when running imagr, what is the best imsize/cellsize to use? Our notes do specify 8192/0.1, but we are not completely sure whether such large numbers are entirely appropriate.
>
Memory: The 6114 Mbytes is the limit of a 32-bit integer to address a
single array. The pseudo-AP is a single array (of double precision
words) and we use 32-bit integers to address it. To change the code to
use 64-bit integers would be prohibitively expensive. And it would
probably not produce better performance - the algorithms are optimized
fairly well for single-thread code. The best performance win would be
to be able to go multi-thread - again essentially impossible with such a
large base of code.
One thing that is not obvious is that the operating system uses the
memory not used by tasks for the disk IO - it is possible that your data
will fit in RAM (if you have a lot) or that the work file in IMAGR will
fit in RAM. If that is the case, the task will run very fast, if that
is not the case it will be very much slower. So the large memory is not
wasted.
You can make some operations be "parallel" by running more than one task
at the same time - imaging each spectral window in separated copies of
aips (AIPS1, APPIS2, AIPS3, etc). Some powerful machines can manage
this - but be careful not to use the same disk files as input or head
contention will occur.
Your IMAGR parameters are simply WRONG. The cell size and image size
must be set depending on your input data and the size of the region in
the sky that you wish to image. In many cases, with longer wavelength
data anyway, multiple facets are required (at least in AIPS). The task
SETFC will recommend cell size and image size and facet locations. I
use it by first setting CELLSIZE and IMSIZE to 0. After running the
task I delete the output text file and select cell sizes and image sizes
to be "nicer" (0.1 asec is nicer than 0.104598 for example). Then I
re-run it to set the locations of the facets, making a BOXFILE for input
to IMAGR.
Eric Greisen
More information about the Daip
mailing list