Set and get number of threads to be used in data.table functions that are parallelized with OpenMP. The number of threads is initialized when data.table is first loaded in the R session using optional environment variables. Thereafter, the number of threads may be changed by calling setDTthreads. If you change an environment variable using Sys.setenv you will need to call setDTthreads again to reread the environment variables.
setDTthreads(threads = NULL, restore_after_fork = NULL, percent = NULL, throttle = NULL)
getDTthreads(verbose = getOption("datatable.verbose"))A length 1 integer. The old value is returned by setDTthreads so you can store that prior value and pass it to setDTthreads() again after the section of your code where you control the number of threads.
NULL (default) rereads environment variables. 0 means to use all logical CPUs available. Otherwise a number >= 1
Should data.table be multi-threaded after a fork has completed? NULL leaves the current setting unchanged which by default is TRUE. See details below.
If provided it should be a number between 2 and 100; the percentage of logical CPUs to use. By default on startup, 50%.
1024 (default) means that, roughly speaking, a single thread will be used when nrow(DT)<=1024, 2 threads when nrow(DT)<=2048, etc. The throttle is to speed up small data tasks (especially when repeated many times) by not incurring the overhead of managing multiple threads. Hence the number of threads is throttled (restricted) for small tasks.
Display the value of relevant OpenMP settings plus the restore_after_fork internal option.
data.table automatically switches to single threaded mode upon fork (the mechanism used by parallel::mclapply and the foreach package). Otherwise, nested parallelism would very likely overload your CPUs and result in much slower execution. As data.table becomes more parallel internally, we expect explicit user parallelism to be needed less often. The restore_after_fork option controls what happens after the explicit fork parallelism completes. It needs to be at C level so it is not a regular R option using options(). By default data.table will be multi-threaded again; restoring the prior setting of getDTthreads(). But problems have been reported in the past on Mac with Intel OpenMP libraries whereas success has been reported on Linux. If you experience problems after fork, start a new R session and change the default behaviour by calling setDTthreads(restore_after_fork=FALSE) before retrying. Please raise issues on the data.table GitHub issues page.
The number of logical CPUs is determined by the OpenMP function omp_get_num_procs() whose meaning may vary across platforms and OpenMP implementations. setDTthreads() will not allow more than this limit. Neither will it allow more than omp_get_thread_limit() nor the current value of Sys.getenv("OMP_THREAD_LIMIT"). Note that CRAN's daily test system (results for data.table here) sets OMP_THREAD_LIMIT to 2 and should always be respected; e.g., if you have written a package that uses data.table and your package is to be released on CRAN, you should not change OMP_THREAD_LIMIT in your package to a value greater than 2.
Some hardware allows CPUs to be removed and/or replaced while the server is running. If this happens, our understanding is that omp_get_num_procs() will reflect the new number of processors available. But if this happens after data.table started, setDTthreads(...) will need to be called again by you before data.table will reflect the change. If you have such hardware, please let us know your experience via GitHub issues / feature requests.
Use getDTthreads(verbose=TRUE) to see the relevant environment variables, their values and the current number of threads data.table is using. For example, the environment variable R_DATATABLE_NUM_PROCS_PERCENT can be used to change the default number of logical CPUs from 50% to another value between 2 and 100. If you change these environment variables using `Sys.setenv()` after data.table and/or OpenMP has initialized then you will need to call setDTthreads(threads=NULL) to reread their current values. getDTthreads() merely retrieves the internal value that was set by the last call to setDTthreads(). setDTthreads(threads=NULL) is called when data.table is first loaded and is not called again unless you call it.
setDTthreads() affects data.table only and does not change R itself or other packages using OpenMP. We have followed the advice of section 1.2.1.1 in the R-exts manual: "... or, better, for the regions in your code as part of their specification... num_threads(nthreads)... That way you only control your own code and not that of other OpenMP users." Every parallel region in data.table contain a num_threads(getDTthreads()) directive. This is mandated by a grep in data.table's quality control script.
setDTthreads(0) is the same as setDTthreads(percent=100); i.e. use all logical CPUs, subject to Sys.getenv("OMP_THREAD_LIMIT"). Please note again that CRAN's daily test system sets OMP_THREAD_LIMIT to 2, so developers of CRAN packages should never change OMP_THREAD_LIMIT inside their package to a value greater than 2.
Internally parallelized code is used in the following places:
between.c - between()
cj.c - CJ()
coalesce.c - fcoalesce()
fifelse.c - fifelse()
fread.c - fread()
forder.c, fsort.c, and reorder.c - forder() and related
froll.c, frolladaptive.c, and frollR.c - froll() and family
fwrite.c - fwrite()
gsumm.c - GForce in various places, see GForce
nafill.c - nafill()
subset.c - Used in [.data.table subsetting
types.c - Internal testing usage