Benchmark code comparing the performance of .combine = "rbind" vs. manual aggregation using data.table::rbindlist() in parallel foreach loops in R. Demonstrates significant speed improvements with manual aggregation.

foreach-rbind-benchmark.R

library(doFuture)

plan(multisession, workers = 32)

options(scipen = 999)

anysvd <- function(id, dim = 10, nrep = 300) {
  results <- vector("list", nrep)

  for (j in 1:nrep) {
    X <- matrix(rnorm(dim^2), dim, dim)
    A <- crossprod(X)

    s <- svd(A)

    results[[j]] <- data.frame(
      id = id,
      sub_id = j,
      sv1 = s$d[1],
      sv2 = s$d[2],
      sv3 = s$d[3],
      matrix_norm = norm(A, type = "F"),
      matrix_trace = sum(diag(A))
    )
  }

  as.data.frame(data.table::rbindlist(results))
}

nsim_grid <- c(1000, 10000, 50000, 100000)

df_bench <- data.frame(
  nsim = integer(),
  method = character(),
  time = numeric()
)

for (nsim in nsim_grid) {
  message("Running benchmark with nsim=", nsim)

  # Method 1: foreach(.combine = "rbind")
  set.seed(42)
  tictoc::tic.clearlog()
  tictoc::tic(paste0("method1_nsim", nsim))
  df_rbind <- foreach(
    i = 1:nsim,
    .combine = "rbind",
    .options.future = list(seed = TRUE)
  ) %dofuture% {
    anysvd(i)
  }
  tictoc::toc(log = TRUE, quiet = TRUE)

  # Method 2: foreach then rbindlist
  set.seed(42)
  tictoc::tic(paste0("method2_nsim", nsim))
  lst_rbindlist <- foreach(
    i = 1:nsim,
    .options.future = list(seed = TRUE)
  ) %dofuture% {
    anysvd(i)
  }
  df_rbindlist <- as.data.frame(data.table::rbindlist(lst_rbindlist))
  tictoc::toc(log = TRUE, quiet = TRUE)

  if (!identical(df_rbind, df_rbindlist)) warning("Discrepant results for nsim=", nsim)

  lst_log <- tictoc::tic.log(format = FALSE)
  df_bench <- rbind(
    df_bench,
    data.frame(
      nsim = nsim,
      method = "foreach(.combine=\"rbind\")",
      time = lst_log[[1]]$toc - lst_log[[1]]$tic
    ),
    data.frame(
      nsim = nsim,
      method = "rbindlist",
      time = lst_log[[2]]$toc - lst_log[[2]]$tic
    ),
    make.row.names = FALSE
  )

  tictoc::tic.clearlog()
}

knitr::kable(df_bench)

ragg::agg_png("foreach-rbind-benchmark.png", width = 2000, height = 2000 / 1.618, res = 300, scaling = 1.25)
ggplot2::ggplot(
  df_bench,
  ggplot2::aes(x = nsim, y = time, color = method, group = method)
) +
  ggplot2::geom_line(linewidth = 1) +
  ggplot2::geom_point(size = 3) +
  ggplot2::labs(
    title = "Speed Comparison",
    subtitle = "foreach(.combine=\"rbind\") vs. rbindlist",
    x = "Number of Simulations (nsim)",
    y = "Time (seconds)",
    color = "Method"
  ) +
  cowplot::theme_cowplot() +
  ggsci::scale_color_observable() +
  ggplot2::guides(colour = ggplot2::guide_legend(position = "inside")) +
  ggplot2::theme(legend.position.inside = c(0.05, 0.8))
dev.off()

options(scipen = 0)