is this scaling acceptable?

[Mohammad Safarzadeh]

I ran mcrx run with 13 cameras and  Smodel.fits for mappings_sed_file / it took about 17 hours to finish the run, with ir_tol = 0.9! output file size about 140GB.

I ran mcrx run with  2  cameras and Smodel-lores128.fits for  mappings_sed_file, it took about 6 hours. for ir_tol=0.9  , output file size is about 3GB.

so there is 3 times more speed gained by changing the input sed model!

If most of the time is being spent on saving data to files, I think camera_aux is something un-necessary, I don't know if anyone uses those infos and if one can set the sims not to save that stage and if that can make it faster.

what is really the accelerating factor here? cameras or the low res sed file?

and also I wonder if using low res files are just giving low res results and the physics of the out put is not corrupted(meaning the result id low res version of the high res result and physics is not lost using them)

Best,

Mohammad

[Patrik Jonsson]

Scaling depends on which calculation you are doing. There's some info
on this in the first Sunrise paper.

When you are shooting rays to cameras (ie not IR equilibrium) then
runtime becomes proportional to number of cameras at some fairly small
number of cameras.

During IR equilibrium, cameras aren't used so they are not a factor there.

During all phases of the calculation, runtime is very close to
proportional to number of wavelengths, because the runtime is
dominated by the vector operations of calculating the polychromatic
bias factors over the wavelength vectors.

Output file size (AUX images take a negligible amount of space and
time) is proportional to number of pixels * number of cameras * number
of wavelengths.

Using lower resolution spectra means that you will miss any features
that are within the bins. Initial luminosity is close to conserved and
opacity is averaged within the bin, but averaging opacity and doing
the calculation is not the same as doing the calculation and then
averaging the results, exponentials don't average like that. So, yes,
physics is "lost" and there is a bias resulting from lower resolution.
The magnitude depends on the opacity curve. You of course also miss
the ability to look at line fluxes with the low res spectra. Only you
can decide if the trade off is acceptable.

cheers,

/Patrik

> --
> You received this message because you are subscribed to the Google Groups
> "Sunrise" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to sunrisemcrx...@googlegroups.com.
> To post to this group, send email to sunri...@googlegroups.com.
> Visit this group at http://groups.google.com/group/sunrisemcrx?hl=en.
> For more options, visit https://groups.google.com/groups/opt_out.
>
>

[Chris Hayward]

To add to what Patrik said: to see a concrete demonstration of how to check whether using lower resolution is okay for your topic, I recommend checking out Fig. 2 of Snyder et al. (2012; http://arxiv.org/abs/1210.6347v1) and the relevant discussion. There, Greg demonstrates that the mid-IR AGN indicators in which he is interested are essentially unaffected by the transition from the full 963 wavelengths to the lower-resolution SEDs with only 128 wavelengths. However, this conclusion is completely problem-dependent, and there are some problems (e.g., doing kinematics) for which you must have higher resolution than that offered by the standard Smodel.fits.

Cheers,

Chris

--
Chris Hayward
Heidelberger Institut für Theoretische Studien
Schloss-Wolfsbrunnenweg 35
69118 Heidelberg, Germany
Google Voice: +1 (617) 744-9416
Office: +49 6221 533 284
Fax: +49 6221 533 298
http://www.cfa.harvard.edu/~chayward