Hi Jure,

I have experienced similar issues in the past. I think this depends on the
R library you use for parallelization but I don't remember the details.

I suggest that you also define OPENBLAS_NUM_THREADS with the same value as
OMP_NUM_THREADS. Defining both env vars with the same value shouldn't hurt
and In theory openblas won't respect the value of OMP_NUM_THREADS unless it
has been compiled with USE_OPENMP=1 which I think it's not the case in
foss/2016b (see
https://github.com/hpcugent/easybuild-easyconfigs/blob/master/easybuild/easyconfigs/o/OpenBLAS/OpenBLAS-0.2.18-GCC-5.4.0-2.26-LAPACK-3.6.1.eb
)

Also see here https://github.com/xianyi/OpenBLAS#usages

regards,
Pablo.

2017-02-02 13:20 GMT+01:00 Jure Pečar <[email protected]>:

>
> Hi all,
>
> I have here R-3.3.1 built with foss-2016b. Ganglia showed me lots of red
> on the nodes where this R is running. I found out that by default this R
> will spawn number of threads equal to number of cpus on the node (does this
> come from openblas?), irregardless of what slurm is limiting the job to.
> So I prepared some prolog scripts that set OMP_NUM_THREADS based on
> --cpus-per-task and did some testing with http://r.research.att.com/
> benchmarks/R-benchmark-25.R
> These are the numbers:
> cores
> 1       37.77user 0.93system 0:39.49elapsed 98%CPU
> 2       39.70user 5.83system 0:36.90elapsed 123%CPU
> 4       42.12user 12.13system 0:34.71elapsed 156%CPU
> 8       46.62user 25.82system 0:33.80elapsed 214%CPU
> 16      52.83user 53.06system 0:33.29elapsed 318%CPU
> Trend is obvious ...
>
> I want to know is this is expected R behaviour, since my experience with
> it is limited. Also my google-fu for single letter things is limited, so I
> would appreciate any pointers to articles I have to read or experience you
> can share. Thanks.
>
>
> --
>
>   Jurij Pečar
>   HPC Engineer, IT Operations, IT Services
>   EMBL Heidelberg, Meyerhofstraße 1, 69117, Heidelberg, Germany
>   Room 13-401
>



-- 
Pablo Escobar López
HPC systems engineer
sciCORE, University of Basel
SIB Swiss Institute of Bioinformatics
http://scicore.unibas.ch

Reply via email to