Time for standard convergence on the full Planck likelihood (TTTEEE)?

[tsmith2]

I running PolyChord on the full Planck likelihood (TTTEEE), and I am wondering if anyone can give me a sense of how long it should take to reach convergence (with the 'standard' parameters from Cobaya: nlive: 25d and precision_criterion: 0.001) for LCDM using 20 MPI processes (2 cores each process).

[lukashergt]

Hi @tsmith2

Assuming the following setup:

• running with Cobaya
• theory code CLASS version 2.9.4
• nonlinear code halofit
• cosmology is standard LCDM
• Planck likelihoods TTTEEE+lowl+lowE
• sampler PolyChord version 1.18.1
• ndim=6+1+20=27 sampling parameters and thus 25*27=675 live points
• num_repeats=2d with speed blocking of 1, 15, 15 and thus the PolyChord number of repeats are 12, 30, 600
• 40 computing cores (Not sure what you mean by "2 cores each process", you should not be using OpenMP when running with PolyChord...)

In this case I'd expect a computing time of about 5d 5h = 125h. Note that this is based on a very rough projection from my runs with different nlive and ncores.

---

Note:

• I don't recommend doing your first run with nlive: 25d! Start out with fewer live points, check that things look sensible, and only then set up the big runs with large number of live points.
• Using CAMB instead of CLASS seems to take half as long (no clue why... would love to find out).
• Using hmcode instead of halofit will take longer.