The compiled code shares its larger jobs between threads: the distances
between rows, the searches for the nearest cluster and the distance
updates of the linkage while many clusters remain (more than about 1,500),
the groups and the nearest-medoid search of two-stage clustering, the
resamples of branch support (tree_support_x,
tree_support_y) and the larger steps and blocks of optimal leaf
ordering. Results are the same whatever the number of threads.
Details
By default HeatmapR uses one thread per physical core, but no more than the
CPUs the R process may run on (on Linux, e.g. when limited by
taskset or a container), no more than the environment variables
OMP_THREAD_LIMIT and OMP_NUM_THREADS allow (when they are
set) and no more than two while R CMD check limits the number of
cores (_R_CHECK_LIMIT_CORES_). Processes forked from R (e.g. by
parallel::mclapply()) use one thread unless the number is set. The
linkage stops sharing its scans between threads when waiting for them
takes longer than the work they save, e.g. when other processes keep the
cores busy; results are the same either way. Set the number
with heat_map_threads(n) or options(HeatmapR.threads = n),
e.g. heat_map_threads(1) when heatmaps are drawn in parallel
processes.
