GNU R vs Julia: is it only a matter of devectorization?

Recently I have read a post on benefits of code devectorization in Julia. The examples given there inspired me to perform my own devectorization exercise. I decided to use bootstrapping as a test ground. The results are quite interesting (and not so bad for GNU R).

The task is very simple (and typical):

1. generate 10000 elements sample from uniform distribution
2. 1000 times perform bootstrap sample of the vector and calculate its standard deviation
3. return standard deviation of the bootstrap distribution
4. Perform steps 1-3 four times and record computation time

Let us start with GNU R implementation:
run  <- function() {
    ssize <- 10000
    nboot <- 1000
    x <- runif(ssize)
    y <- replicate(nboot, sd(sample(x, ssize, TRUE)))
    sd(y)
}

for (i in 1:4) {
    cat(system.time(run())[3], ” “)
}
# result: 0.34  0.32  0.31  0.34
The direct translation to Julia gives:
using Distributions

function run()
    ssize = 10000
    nboot = 1000
    x = rand(ssize)
    y = Array(Float64,nboot)
    for i in 1:nboot
        y[i] = std(sample(x, ssize))
    end
    std(y)
end

for i in 1:4
    print(“$(@elapsed run())  “)
end
# result: 0.38965987  0.331032088  0.3208918  0.315452803