a performance benchmark for https://github.com/Rdatatable/data.table/pull/1613

fwrite-performance-test.R

library(data.table)
library(microbenchmark)

performanceTest <- function (nrows=100000, ncols=400, quote=TRUE, col.type='character') {
  
  message(paste0('generating test table, col type ', col.type))
  x <- c(1:nrows)
  dt <- data.table(col1=x)
  
  col_generators = list(
    character = function(x) paste0('char', x),
    double = function(x) as.double(x / 0.9),
    integer = function(x) x,
    factor = function(x) as.factor(paste0('factor', x %% 1000))
  )
  
  if (col.type != 'mixed') {
    col_generators <- col_generators[col.type]
  }
  
  for (col in c(1:ncols)) {
    col_generator <- col_generators[[((col-1) %% length(col_generators))+1]]
    dt[, (paste0('col', col)) := col_generator(x)]
  }
  
  message(paste0('dt of size ', nrow(dt), 'x', ncol(dt), ' quoting ', quote))
  
  timeIt <- function(func) {
    f <- tempfile()
    t0 <- get_nanotime()
    func(dt, f)
    time_diff <- (get_nanotime() - t0) * 1e-9
    file.remove(f)
    time_diff
  }
  
  alternatives = list(
    fwrite = function(dt, f) {
      fwrite(dt, f, quote=quote)
    },
    write.csv = function(dt, f) {
      write.csv(dt, f, quote=quote, row.names=FALSE)
    }
  )
  
  n_rounds <- 3
  
  results <- rbindlist(
    lapply(1:n_rounds,
       function (i) {
         round <- list()
         for (alt in names(alternatives)) {
           message(alt, ' ', i)
           round[[alt]] <- timeIt(alternatives[[alt]])
         }
         round
       }
    )
  )
  
  means <- results[, lapply(.SD, mean)][, lapply(.SD, as.numeric)]
  fwrite_time <- means[, fwrite]
  speedup <- means[, write.csv] / fwrite_time
  list(speedup=speedup, fwrite_time=fwrite_time)
}

performanceTable <- function() {
  tab <- data.table(expand.grid(list(
      `col.type` = c('character', 'double', 'integer', 'factor', 'mixed'),
      `quote` = c(TRUE, FALSE)
  )))
  
  for (j in 1:nrow(tab)) {
    r <- performanceTest(col.type=tab[j, `col.type`], quote=tab[j, `quote`])
    tab[j, speedup := r$speedup]
    tab[j, fwrite_time := r$fwrite_time]
  }
  
  tab
}

Example output on Debian Linux with R 3.1.1, an SSD, Core-i5 and 8GB RAM.

col.type	quote	speedup	fwrite_time (s)
character	TRUE	1.75357325745381	9.9918577671051
factor	TRUE	2.0468551983038	9.39413905143738
mixed	TRUE	1.34548667345286	22.5989712874095
character	FALSE	1.90443330503114	8.63940691947937
double	FALSE	1.02673089671228	56.5861701965332
integer	FALSE	1.75163767836391	12.2004607518514
factor	FALSE	2.15513619533713	8.37993264198303
mixed	FALSE	1.34038028170995	22.0992030302684

@oseiskar Sys.time() won't time nicely, for a reason why see gc argument ?system.time.

Ran again with this version:

col.type	quote	speedup	fwrite_time (s)
character	TRUE	2.03	8.941
double	TRUE	3.27	17.885
integer	TRUE	4.33	4.943
factor	TRUE	2.31	8.409
mixed	TRUE	2.99	10.181
character	FALSE	2.15	7.594
factor	FALSE	2.33	7.643
mixed	FALSE	3.027	9.749

Updated code to use microbenchmark::get_nanotime() (thanks @MichaelChirico ).
Results are similar.

col.type	quote	speedup	fwrite_time (s)
character	TRUE	2.00	8.91
double	TRUE	3.25	17.85
integer	TRUE	4.35	4.92
factor	TRUE	2.21	8.5
mixed	TRUE	2.91	10.31
character	FALSE	2.17	7.57
factor	FALSE	2.30	7.75
mixed	FALSE	3.03	9.68