Apr. 2013 CosmoMC: problem thinning long chains - crashes in FreeItem

[Geraint Harker]

I wondered if anyone else had had a problem with long chains in the current version of CosmoMC. Once there have been 500000 steps (by default), S%Thin is called from AddMPIParams. When this method tries to free up the samples that have been thinned out, I get the following error, which according to the traceback occurs in the following line of FreeItem:

if (L%OwnsObjects .and. associated(Point%P)) deallocate(Point%P)

The error message I get is as follows:

...
Chain 1 MPI communicating
Current convergence R-1 = 0.1326408 chain steps = 500501
slow changes 0 power changes 0
forrtl: severe (173): A pointer passed to DEALLOCATE points to an object that cannot be deallocated
Image PC Routine Line Source
cosmomc 00000000006BFA6E Unknown Unknown Unknown
cosmomc 00000000006BE506 Unknown Unknown Unknown
cosmomc 0000000000676C72 Unknown Unknown Unknown
cosmomc 0000000000600FBC Unknown Unknown Unknown
cosmomc 00000000005F4196 Unknown Unknown Unknown
cosmomc 0000000000422658 objectlists_mp_fr 172 ObjectLists.f90
cosmomc 0000000000424CEF objectlists_mp_th 359 ObjectLists.f90
...

I can't see what the problem is, though I haven't completely grokked all the Fortran 2003 features yet. I'm using Intel 13.0.0 and OpenMPI 1.6.4 on RHEL5.

Thanks in advance for any help!

[Antony Lewis]

I can't replicate an issue on slightly higher ifort version test case, but let me know if you find a fix.

(do you really need so many samples!?)

[Geraint Harker]

OK, I'll keep working on it. I'm running in quite a few dimensions (68 for this problem) and it seems to be taking this many samples to converge. I guess I could avoid the issue by computing the covariance matrix and using that as the initial guess in a new run. They're installing redhat 6 on Wednesday, possibly with a new version of the compilers, I'll see if the issue persists then.

Thanks!
Geraint