cosmoMC output + Compilation error

Use of Healpix, camb, CLASS, cosmomc, compilers, etc.
Post Reply
Rita Sinha
Posts: 3
Joined: September 29 2004
Affiliation: IUCAA, Pune, India

cosmoMC output + Compilation error

Post by Rita Sinha » August 31 2005

Hi,
I have 2 queries one regarding cosmoMC output and the other one a compiler issue:

1)This is regarding the output files after doing 'getdist'. As I understand the likestats file contains the best-fit values and their -sigma and 2-sigma confidence intervals. The margestats file contains the mean values, standard deviations and the marginalized limits for the different parameters. It is known that (also mentioned in the cosmomc readme file) that MCMC does not provide much accurate value of the best fit values. How do I get reasonable error bars on my best fit parameters obtained by MCMC. I am unable to figure out. Please suggest.

2) We have a 64 bit 8x4 AlphaServer ES45 68/1250 System (@ 1.25 GHz). I have been running cosmoMC on this cluster for some time now using HP Fortran 5.4 compiler.
Recently we upgraged to HP Fortran 5.5A compiler and cosmoMC gave compilation errors. Since we were unable to resolve this issue we downgraded to Compaq Fortran V5.5-1877 and everything worked as usual. In general, I found out from HP Fortran website that V5.5A has the following known problem:
"Version 5.5A Known Problems
* Fortran note 3264 - The compiler dies at compile-time with the message
"**Internal compiler error: internal abort**"
in some codes that use MODULEs with PUBLIC and PRIVATE data."

The errors which I encountered while compiling cosmoMC and some tentative solutions which I tried are listed below:

Errors with CAMB directory:
---------------------------------
1)***Internal compiler error: internal abort**
Please report this error along with the circumstances in which it occurred in a Software Problem Report. Note: File and line given may not be explicit cause of this error.
!$ if (ThreadNum /=0) call OMP_SET_NUM_THREADS(ThreadNum)

Solution: commenting out this line compiles CAMB sucessfully.
(alternatively one can give 'setenv OMP_NUM_THREADS 4' on command line)

Errors with source directory:
----------------------------------
1) **Internal compiler error: internal abort** Please report this error along with the circumstances in which it occurred in a Software Problem Report. Note:File and line given may not be explicit cause of this error.
Threadnum =num_threads

Solution: commenting out this line omits this particular error but compilation error again at another place.

2) Subsequent compiling action in the end gives the error:
**Internal compiler error: internal abort** Please report this error along with the
circumstances in which it occurred in a Software Problem Report. Note: File and line given may not be explicit cause of this error.

In the end incomplete compilation of 4 files (CMB_Cls_simple.f90, MCMC.f90, calclike.f90, paramdef.f90) is reported although these empty files are created.

One can possibly remove "private" attribute in modules by the use of (for example):
'USE' module_a, ONLY: param1
or wherever applicable
by renaming the earlier used procedure/variable such as:
USE module, param1 => cparam1

These are sometimes not feasible and cumbersome. Also we are not sure that this will remove the errors with HP Fortran V5.5A. In case anyone has been using this compiler
or has across a similar error please give the fix or suggest if something more sensible can be done.

Thanks in advance.

Regards
Rita Sinha

Antony Lewis
Posts: 1512
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: cosmoMC output + Compilation error

Post by Antony Lewis » August 31 2005

Rita Sinha wrote: 1)This is regarding the output files after doing 'getdist'. As I understand the likestats file contains the best-fit values and their -sigma and 2-sigma confidence intervals. The margestats file contains the mean values, standard deviations and the marginalized limits for the different parameters. It is known that (also mentioned in the cosmomc readme file) that MCMC does not provide much accurate value of the best fit values. How do I get reasonable error bars on my best fit parameters obtained by MCMC. I am unable to figure out. Please suggest.
The .likestats file containts the parameters of the best fit sample, together with the parameter extrema corresponding to the N-D limits (see the CosmoMC paper for discussion of different types of limit).

Usually people quote the less conservative marginalized limits from the .margestats file. You could combine the best-fit value with error bars from the .margestats file, or quote the marginalized means.

I'm not aware of any easy way to immediately quantify the likely sampling error on the parameters of the best-fit sample. Try running with action=2 if you want to get a more accurate value of the best fit point using conjugate gradients.

Post Reply