Performance question

[Russell-Ryan]

I'm following the examples to create a spectrum, but it takes ~30 s on my modern apple laptop. I was curious if this performance is expected, or did I build the code incorrectly somehow?

My laptop is an Apple M1 Max with 64 Gb of memory.

[bd-j]

Hi @Russell-Ryan. Yes, depending on the isochrone library, the initial SSP generation for a given metallicity can be time consuming. However, each SSP is cached so as long as none of the fsps.ParameterSet.ssp_params are changed, subsequent calls for a given metallicity should be much faster.

[Russell-Ryan]

@bd-j thanks! I was finding that, but for each call I will need to adjust the metallicity, stellar mass, SFH, and dust attenuation. I found that the BaSeL library was much faster than the MILES, I presume because of the fewer wavelength steps. But we're not looking for extremely precise results out of this, we're trying to make some vaguely realistic spectra for some simulations for Roman. So do you have any tips for which ISOCHRONE or maybe DUST library might get us the best performance?

[bd-j]

PADOVA isochrones will probably be a somewhat faster.

Just to make sure, you are doing something like

# Instantiate once
sp = fsps.StellarPopulation()
# adjust parameters and get spectrum
for dust in [0.0, 0.5, 1.0]:
    sp.params["dust2"] = dust
    wave, spec = sp.get_spectrum(tage=tage)

And not

for dust in [0.0, 0.5, 1.0]:
    # Instantiating in the loop, very slow, all cached info erased each time!
    sp = fsps.StellarPopulation(dust2=dust)
    wave, spec = sp.get_spectrum(tage=tage)