# ========================
library(microbenchmark)
library(compiler)
library(plyr)
square <- function(x) x^2
f_for <- function(n) for (x in 1:n) x^2
f_lapply <- function(n) lapply(1:n, square)
f_vapply <- function(n) vapply(1:n, square, numeric(1))
c_square <- cmpfun(square)
c_for <- cmpfun(f_for)
c_lapply <- cmpfun(function(n) lapply(1:n, c_square))
c_vapply <- cmpfun(function(n) vapply(1:n, c_square, numeric(1)))
# Run benchmarks for n=10^pow10 cycles
run_benchmarks <- function(pow10) {
n <- 10^pow10
# Run a maximum of 100 iterations, and drop off until there's 1 iteration
# for 1e7
times <- round(min(1e+07/n, 100))
res <- microbenchmark(f_for(n), f_lapply(n), f_vapply(n), c_for(n), c_lapply(n),
c_vapply(n), times = times)
res$size <- n
res
}
# Run benchmarks
res <- ldply(1:7, run_benchmarks)
# Summarize
res_summary <- ddply(res, c("expr", "size"), summarise, N = length(time), m_time = mean(time),
sd = sd(time), se = sd/sqrt(N))
# Calculate time per iteration
res_summary$iter_time <- res_summary$m_time/res_summary$size
# Put compiled and function type into separate columns
res_summary$compiled <- "non-compiled"
res_summary$compiled[grepl("^c_", res_summary$expr)] <- "compiled"
res_summary$type <- sub("^._", "", res_summary$expr)
# print it out
arrange(res_summary, size, compiled, type)
## expr size N m_time sd se iter_time
## 1 c_for(n) 1e+01 100 2.331e+03 6.056e+02 6.056e+01 233.1
## 2 c_lapply(n) 1e+01 100 1.012e+04 2.071e+03 2.071e+02 1012.1
## 3 c_vapply(n) 1e+01 100 1.184e+04 4.055e+03 4.055e+02 1183.9
## 4 f_for(n) 1e+01 100 3.411e+03 5.213e+02 5.213e+01 341.1
## 5 f_lapply(n) 1e+01 100 1.165e+04 3.163e+03 3.163e+02 1164.8
## 6 f_vapply(n) 1e+01 100 1.306e+04 2.648e+03 2.648e+02 1306.0
## 7 c_for(n) 1e+02 100 1.990e+04 6.525e+04 6.525e+03 199.0
## 8 c_lapply(n) 1e+02 100 7.398e+04 1.308e+04 1.308e+03 739.8
## 9 c_vapply(n) 1e+02 100 7.766e+04 1.436e+04 1.436e+03 776.6
## 10 f_for(n) 1e+02 100 3.209e+04 6.639e+04 6.639e+03 320.9
## 11 f_lapply(n) 1e+02 100 8.805e+04 1.084e+04 1.084e+03 880.5
## 12 f_vapply(n) 1e+02 100 9.150e+04 1.314e+04 1.314e+03 915.0
## 13 c_for(n) 1e+03 100 1.101e+05 1.598e+04 1.598e+03 110.1
## 14 c_lapply(n) 1e+03 100 6.370e+05 1.282e+05 1.282e+04 637.0
## 15 c_vapply(n) 1e+03 100 7.022e+05 3.589e+05 3.589e+04 702.2
## 16 f_for(n) 1e+03 100 2.291e+05 7.336e+04 7.336e+03 229.1
## 17 f_lapply(n) 1e+03 100 8.094e+05 2.086e+05 2.086e+04 809.4
## 18 f_vapply(n) 1e+03 100 8.147e+05 1.717e+05 1.717e+04 814.7
## 19 c_for(n) 1e+04 100 1.344e+06 4.767e+05 4.767e+04 134.4
## 20 c_lapply(n) 1e+04 100 7.907e+06 3.565e+06 3.565e+05 790.7
## 21 c_vapply(n) 1e+04 100 7.722e+06 2.811e+06 2.811e+05 772.2
## 22 f_for(n) 1e+04 100 2.452e+06 4.524e+05 4.524e+04 245.2
## 23 f_lapply(n) 1e+04 100 8.692e+06 1.407e+06 1.407e+05 869.2
## 24 f_vapply(n) 1e+04 100 8.749e+06 9.409e+05 9.409e+04 874.9
## 25 c_for(n) 1e+05 100 1.485e+07 7.084e+06 7.084e+05 148.5
## 26 c_lapply(n) 1e+05 100 9.667e+07 2.086e+07 2.086e+06 966.7
## 27 c_vapply(n) 1e+05 100 7.923e+07 1.625e+07 1.625e+06 792.3
## 28 f_for(n) 1e+05 100 2.661e+07 7.814e+06 7.814e+05 266.1
## 29 f_lapply(n) 1e+05 100 1.062e+08 2.215e+07 2.215e+06 1061.8
## 30 f_vapply(n) 1e+05 100 9.298e+07 1.741e+07 1.741e+06 929.8
## 31 c_for(n) 1e+06 10 1.663e+08 6.281e+07 1.986e+07 166.3
## 32 c_lapply(n) 1e+06 10 1.777e+09 2.162e+08 6.836e+07 1776.9
## 33 c_vapply(n) 1e+06 10 7.241e+08 7.160e+07 2.264e+07 724.1
## 34 f_for(n) 1e+06 10 2.420e+08 4.454e+07 1.409e+07 242.0
## 35 f_lapply(n) 1e+06 10 1.467e+09 7.177e+07 2.270e+07 1466.9
## 36 f_vapply(n) 1e+06 10 8.320e+08 7.565e+07 2.392e+07 832.0
## 37 c_for(n) 1e+07 1 1.950e+09 NA NA 195.0
## 38 c_lapply(n) 1e+07 1 3.798e+10 NA NA 3798.4
## 39 c_vapply(n) 1e+07 1 6.800e+09 NA NA 680.0
## 40 f_for(n) 1e+07 1 2.266e+09 NA NA 226.6
## 41 f_lapply(n) 1e+07 1 2.152e+10 NA NA 2152.3
## 42 f_vapply(n) 1e+07 1 8.894e+09 NA NA 889.4
## compiled type
## 1 compiled for(n)
## 2 compiled lapply(n)
## 3 compiled vapply(n)
## 4 non-compiled for(n)
## 5 non-compiled lapply(n)
## 6 non-compiled vapply(n)
## 7 compiled for(n)
## 8 compiled lapply(n)
## 9 compiled vapply(n)
## 10 non-compiled for(n)
## 11 non-compiled lapply(n)
## 12 non-compiled vapply(n)
## 13 compiled for(n)
## 14 compiled lapply(n)
## 15 compiled vapply(n)
## 16 non-compiled for(n)
## 17 non-compiled lapply(n)
## 18 non-compiled vapply(n)
## 19 compiled for(n)
## 20 compiled lapply(n)
## 21 compiled vapply(n)
## 22 non-compiled for(n)
## 23 non-compiled lapply(n)
## 24 non-compiled vapply(n)
## 25 compiled for(n)
## 26 compiled lapply(n)
## 27 compiled vapply(n)
## 28 non-compiled for(n)
## 29 non-compiled lapply(n)
## 30 non-compiled vapply(n)
## 31 compiled for(n)
## 32 compiled lapply(n)
## 33 compiled vapply(n)
## 34 non-compiled for(n)
## 35 non-compiled lapply(n)
## 36 non-compiled vapply(n)
## 37 compiled for(n)
## 38 compiled lapply(n)
## 39 compiled vapply(n)
## 40 non-compiled for(n)
## 41 non-compiled lapply(n)
## 42 non-compiled vapply(n)
# Plot
library(ggplot2)
library(scales)
xlog <- scale_x_log10(breaks = trans_breaks("log10", function(x) 10^x), labels = trans_format("log10",
math_format(10^.x)))
ylog <- scale_y_log10(breaks = trans_breaks("log10", function(x) 10^x), labels = trans_format("log10",
math_format(10^.x)))
p <- ggplot(res_summary, aes(x = size, y = m_time, linetype = compiled, colour = type)) +
geom_line() + geom_point()
# Show with linear scale: large values are more expensive per cycle
p + xlab("size of data") + ylab("total time (ns)")
# Show with log-log scale
p + xlog + ylog + xlab("size of data (log)") + ylab("total time (ns) (log)")
# Time per iteration on y, log(size) on x
ggplot(res_summary, aes(x = size, y = iter_time, linetype = compiled, colour = type)) +
geom_line() + geom_point() + xlog + xlab("size of data (log)") + ylab("Time per iteration (ns)")
