[R-sig-ME] Parallel version of lmer or glmer?

[Douglas Bates]

On Thu, Aug 20, 2015 at 6:35 PM María del Carmen Romero <
mariadelc.romero at gmail.com> wrote:

> Hello,
> I want to know if there is a parallel version of lmer or glmer (both of
> package lme4).
>

It depends what you mean by "parallel".  Do you mean "faster"? (If instead
you mean "I want to tie up lots of processors for a long time"  just use
MCMC methods. :-)

If you do mean "faster" you can fit LMM or GLMM models faster but exactly
how to go about it depends very much on the type of problem you have.

- are you fitting LMMs or GLMMs?
- are you using a muilt-core processor or a cluster?
- how many observations, fixed-effects parameters, distinct grouping
factors and covariance parameters?  (If you are fitting a "maximal model"
in the sense of Barr et al. (2012) and it is taking a very long time the
simplest advice is "don't do that".)

Pantelis has mentioned using multi-threaded BLAS such as OpenBLAS or Intel
MKL but without first profiling the execution on your problem it would be
difficult to assess whether this would help.  Most of the numerical linear
algebra in lme4 is performed using Eigen.

For LMMs the simplest approach to making things faster is to change the
optimizer from the default to the version of bobyqa from the nloptr
package. Over the summer Colin Longhurst performed an in-depth comparison
of optimizers on LMM problems.  Preliminary results are at
https://github.com/Stat990-033/Timings/blob/master/inst/doc/Summaries.ipynb

	[[alternative HTML version deleted]]