MPI question - cmdstan.2.18.1


Well, after switching to gcc 5.3.0, gcc openmpi, and cloning cmdstan from git it started working.


After installing cmdstan.2.8.11 from tar,
adding to make/local
make build-mpi
make build

and making the stan model and running it with mpirun -n 2 (intel MPI) the code runs but you can clearly see that different chains run on different cores and output of those chains is combined. The same code for cmdstan.2.18.0 and on different cluster ran fine and output indicated there was a single chain.

uname -a gives
Linux 3.10.0-693.21.1.el7.x86_64 #1 SMP Wed Mar 7 19:03:37 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux

g++ -v gives
gcc version 4.8.5 20150623 (Red Hat 4.8.5-16) (GCC)

stan model contains in transformed parameters block - I just needed to output log likelihood
lp = sum(map_rect(bnn, phi, theta, x_r, x_i));
and in model block
target += lp;

Any advice?


Yeah, the makefiles for MPI in 2.18.1 are broken and need a small fix which I forgot where I put it. The develop version is working nicely thanks to @syclik’s refactored makefiles. Sorry for that.