Skip to contents

Configure Parallel Execution Backends

Source: R/03-setThreads.R
setThreads.Rd

Unified interface to configure thread counts and acceleration features across system-level (OpenMP/data.table/cheapr) and TensorFlow backends. Uses a hierarchical configuration model where global threads serves as default for all backends unless explicitly overridden.

Critical: TensorFlow configuration must be set before importing TensorFlow via reticulate::import("tensorflow").

Usage

setThreads(
  threads = NULL,
  backend = c("openmp", "dt", "cheapr"),
  tf_config = list(xla_flag = "--tf_xla_auto_jit=2 --tf_xla_cpu_global_jit", xla_device =
    NULL, inter_op = NULL, intra_op = c(1L, NULL)),
  verbose = getFuncOption("verbose"),
  ...
)

Arguments

threads

Integer. Global thread count used as default for all backends. If NULL (default), uses floor(availableCores() / 2). Applied to: OpenMP, data.table, and TensorFlow intra-op (unless overridden).

backend

Character vector. System-level backends to configure: "openmp" (sets OMP_NUM_THREADS), "dt" (data.table threads). Default: c("openmp", "dt", "cheapr").

tf_config

Named list for TensorFlow-specific configuration:

  • xla_flag: Character. XLA JIT compilation flags (default: auto-optimized)

  • xla_device: Integer. XLA device ID (default: 1L)

  • inter_op: Integer. Inter-op parallelism threads (default: max(2, floor(threads/4)))

  • intra_op: Integer. Intra-op parallelism threads (default: 1L). If NULL, inherits global threads by default.

verbose

Logical. Whether to print verbose output (default: inherited from function options).

...

Additional arguments for data.table::setDTthreads() (e.g., restore).

Value

Invisible list with old/new values per backend.

Examples

if (FALSE) { # \dontrun{
# Basic usage: auto-detect and configure
setThreads()

# Explicit thread count for CPU-intensive workloads
setThreads(threads = 12L)

# TensorFlow-optimized configuration for deep learning
setThreads(
  threads = 8L,
  tf_config = list(
    inter_op = 2L,
    intra_op = 8L
  )
)
library(tensorflow)  # Import AFTER setThreads()

# Configure only data.table for memory-efficient workflows
setThreads(threads = 4, backend = "dt", verbose = FALSE)
} # }