Threading and mpi and tbb and gpu in cmdstan

Interfaces CmdStan
1

Hi,

In order to parallelize sampling I initially used mpi. As I understand as long as you are not using more than 1 node threading is not less effective. Right?

There is a thread Benefits of parallelization with a threadpool of the Intel TBB where it is shown that TBB is more efficient than MPI. Is there any way to invoke different TBB methods as demonstrated in the thread?

I have compiled cmdstan with the following flags in make/local
STAN_OPENCL=true
OPENCL_DEVICE_ID=0
OPENCL_PLATFORM_ID=0

Does it mean that cmdstan for sampling uses GPU exclusively for Cholesky and CPU all other times? Does it use GPU for matrix operations?

Thanks

2

Hi Linas,

If you are using map_rect, which I presume you are since you are using MPI, the only thing you need to do in order to use TBB for threading is:

Regarding OpenCL support:

If you are using Cmdstan 2.21 the OpenCL the following functions will be run on the GPU if the input sizeis large enough:

  • cholesky_decompose
  • matrix mutliplication
  • mdivide_left_tri_low
  • mdivide_right_tri_low
  • gp_cov_exp_quad

The plan is to support most Stan functions with GPUs for the 2.22 release but that doesnt help you here.

3

At the moment the TBB is used to parallelise map_rect and I would expect that MPI will give you the same performance. The thread you refer to doesn’t even compare against MPI as I can see. From my experience, the TBB map_rect is now just as fast as MPI. So threading was lacking in speed behind MPI, but that slowness of threading is now gone with the use of the TBB.

4

Thanks for the help. Should I expect a difference between threading and MPI (in the favor of MPI)?

5

No