Timing estimate - Githubissues

LSSTDESC / CLxTJP

Efforts to test and identify an interface between the CL analyses and FireCrown

BSD 3-Clause "New" or "Revised" License

0 stars 0 forks source link

Timing estimate #1

Open combet opened 4 years ago

combet commented 4 years ago

Following discussions regarding the interfacing of CLMM with Firecrown, the first question that came up was what was the best option for passing tangential shear profiles from CLMM to Firecrown:

"Brute force" (for each point in the sampling in Firecrown call CLMM to provide a g_t profile)
Grid in M and z and then interpolation in Firecrown
Vectorization in CLMM (not implemented yet).

So this issue first aim is to time option 1 and 2.

cavestruz commented 4 years ago

ClusterToolkit does not take an array of masses, and provides DeltaSigma, not tangential shear (which need Sigma_crit(cosmo))
Note, all astropy things do vectorize. SigmaCrit as computed in CLMM can take a single cluster redshift and an array of source redshifts and return array of sigma crits
Need to find out how many calls of the generating model profile we will need for sampling in FireCrown. Right now, all we can do is take 10 masses, and 10 cluster masses, create a grid, and see how long it takes. For each point in MCMC, can generate one of these and later interpolate for FireCrown.

pennalima commented 4 years ago

Hi @combet , @cavestruz and @elikrause,

@vitenti and I made some progress. We obtained that NumCosmo (NC) is faster than cluster-toolkit (CT). The difference between NC and CLMM (wrap CT) is even larger. For example, CT: 2 s ± 107 ms CLMM: 2.58 s ± 89.6 ms NC: 1.51 s ± 40.9 ms when computing excess surface density.

Sandro and I are working on a manner to deal with external libraries without losing velocity. We'll keep you updated.