Running parallel MCMC on multicore workstation

Estimation ML/Bayesian estimation
1

Hi,
I’ve run my MCMC using this command

options_.parallel = struct(‘Local’, 1, ‘Name’,‘n1’,‘CPUnbr’,[1:8],…
‘NumberOfThreadsPerJob’,8,‘NumCPU’, 8, ‘user’,‘’,‘Port’,‘’,‘SingleCompThread’,‘true’,‘passwd’,‘’,‘NodeWeight’,‘1’, ‘RemoteDrive’, ‘’,…
‘RemoteFolder’,‘’,‘RemoteDirectory’,‘’, ‘MatlabOctavePath’, ‘matlab’, ‘DynarePath’, ‘’,‘UserName’,‘gianni’,‘ComputerName’,‘localhost’);

estimation(datafile=‘US_DNGS15_data.mat’,mode_compute=0,mode_file=US_DNGS15_mh_mode,nobs=230,presample=40,mh_replic=1000,mh_jscale=0.20,mh_nblocks=6)
DL_GDPC,
DL_GDPC,
DL_GDPDEF,
DL_PCEC_GDP,
DL_FPI_GDP,
DL_AWHNONAG,
DL_COMPNFB,
FFR;

Yet the MCMC seems to be run sequentially (one chain after the other).
Does anybody know what am I missing?
Cheers
Gianni

2

@lombard

  1. Which version of Dynare are you using?
  2. Please set the OperatingSystem option and fill out the DynarePath
  3. Did you run the parallel_test?
  4. Do you get any error message?
  5. Which operating system are you on?

@rattoma Do you have any insights? I found a bug in 4.5.3 (https://github.com/DynareTeam/dynare/issues/1548), but even after that, I get a crash, because Parallel(indPC).OperatingSystem is not set in dynareParallelDelete.m, but according to the manual, one does not need to set it.

3

Thanks Johannes,
I made it work by setting up a config file and dropping the option line.
I’m still using a development version of Dynare (16-03-2017), I’m on a Fedora Linux workstation and have to
recompile dynare from source… Plus I made few changes to the estimation file of Dynare (see my other post) so changing it is a bit costly.
… but now it works!
Cheers
Gianni

4

@lombard
Hi Gianni,
could you then please provide me with the content of options_.parallel once the run is done. Comparing the field resulting from the config file to the one derived from the option line should allow for debugging.