Skip to contents

Cluster computation

Cluster.Rmd

To speed up the computation, the batss.glm function can use parallelisation on single machines when computation = "parallel" (which is the default).

When using a cluster, parallelisation is best achieved by letting the cluster workload manager - typically Slurm on clusters running Linux - split the set of seeds (corresponding to as many simulated trials) between cluster nodes and cpus.

Let’s assume a BATSS user wants to perform a Monte Carlo simulation considering 10’000 trials and has 500 cpus to do so. The strategy we suggest consists in

  • running batss.glm on each cpu with a subset of the 10’000 seeds of interest specified in argument R, so that each cpu evaluates a different set of seeds,
  • saving the (500) batss.glm outputs as a Rdata files with the function save under different names (like one of the seed evaluated by the cpu like the first or the last one, for example),
  • finally use the function batss.combine to combine these outputs.

In the next Section, we show examples of use of the function batss.combine.