limit BLAS/MKL-based parallelism in DE methods

[grst]

Description of the bug

There's two ways of parallelization in DE methods:

• implicit paralellism via the underlying linear algebra libraries (applied during operations such as matrix multiplications)
• explicit paralellism via Bioc parallel.

If both of them are active, this leads to a massive overbooking of threads, and therefore a high overhead cost, effectively slowing down the analysis.

Since the explicit parallelism is very effective for DE methods (it's possible to fit models for different genes in parallel) I'd recommend to disable the implicit parellelism althogether. This can be done with the RhpcBLASctl package:

RhpcBLASctl::blas_set_num_threads(1)
RhpcBLASctl::omp_set_num_threads(1)

Command used and terminal output

Relevant files

No response

System information

No response

[grst]

This is especially relevant for DREAM, as it is much more computationally expensive than the other methods.
I suggest to do some testing with a complex formula including a mixed effect. Then try

Biocparallel threads ->
↓ BLAS/MKL trheads

1	2	4	8	16
2
4
8
16

and note for each configuration the wall time, cpu time, and peak cpu usage.

We could then settle on the configuration that is fastest/most efficient.