Baseline run not converging

Use of Cobaya. camb, CLASS, cosmomc, compilers, etc.
Post Reply
Danylo Yakymiv
Posts: 4
Joined: July 27 2015
Affiliation: Stony Brook

Baseline run not converging

Post by Danylo Yakymiv » July 27 2015

Hello,

I am attempting to run chains for the baseline set (plik_dx11dr2_HM_v18_TT.ini and lowTEB.ini, as well as the common and calibration ini files) on 48 cores and the chains haven't converged after ~20 hours. Most parameters have converged to gaussians at this point, the only problems being [tex]\xi^{tSZ-CIB}[/tex] and [tex]A^{kSZ}[/tex]. The conditions for convergence in common.ini have not been changed from the defaults. Is this expected and it only needs more time, or is there likely a problem? Thank you.

Jason Dossett
Posts: 97
Joined: March 19 2010
Affiliation: The University of Texas at Dallas
Contact:

Baseline run not converging

Post by Jason Dossett » July 28 2015

How many chains are you actually running? I usually run 6-8 chain jobs with each chain threaded out to 4-8 cores depending on node availability when I start the job. My baseline run took 19 hours running this way.

Antony Lewis
Posts: 1941
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Baseline run not converging

Post by Antony Lewis » July 28 2015

And what convergence settings? The defaults in the July CosmoMC release are actually quite stringent, more than you probably need unless you are then importance sampling or making complex plots (I forgot to change them back to more normal values after the final Planck runs). Usually basic runs will look fine in about 8 hours on a modern cluster using with 4 cores per chain. e.g. you could use

Code: Select all

MPI_Converge_Stop = 0.02
MPI_Check_Limit_Converge = F

Danylo Yakymiv
Posts: 4
Joined: July 27 2015
Affiliation: Stony Brook

Baseline run not converging

Post by Danylo Yakymiv » July 28 2015

I am running 8 chains each on 6 cores, so the timescale matches. I'll try easing up the convergence settings. Thank you for your help.

Post Reply