R vs Python execution time comparisons

Running R could be slow at times and implementing computationally intensive parts in C/C++ is recommended. Writing and debugging C code, however, is time consuming and much less convenient than programming in a scripting language like R or Python. To see how execution times of some simple R code compare to Python implementations, we carry out the following simulations:

* Example 1:

# R code
k <- 0
for(i in 1:10^7)
k <- k+i


# Python code
k = 0
for i in range(10**7):
k = k+i


R execution time (in sec): 15.266
Python execution time (in sec): 4.563




* Example 2:

# R code
for(i in 1:10^5)
for(j in 1:1000)
x <- 1


# Python code
for i in range(10**5):
for j in range(1, 1001):
k = 1


R execution time (in sec): 49.609
Python execution time (in sec): 17.65


* Example 3:

# R code
x <- 1:10^7
for(i in 1:10^7)
k <- x[i]


# Python code
x = range(10**7)
for i in range(10**7):
k = x[i]


R execution time (in sec): 13.563
Python execution time (in sec): 3.563

R speed increase

For those of you who would like to increase computational speed when using R for matrix calculations, consider the following:

For Windows users, R gets shipped with a pre-compiled Rblass.dll which is not optimized for any specific CPU. However, there exist several CPU-specific Rblass.dlls. Try using those!
You can find some here. For my Core2Duo, I had to look here.
Notice that using the right dll could lead to drastically increased computations.

Example (on my Intel Core2Duo 6600@2.4Ghz, 2GB):

Default….

require(Matrix)
set.seed(123)
X <- Matrix(rnorm(1e6), 1000)
> print(system.time(for(i in 1:25) X%*%X))
user system elapsed
43.20 0.18 43.59
> print(system.time(for(i in 1:25) solve(X)))
user system elapsed
24.87 0.03 25.28
> print(system.time(for(i in 1:10) svd(X)))
user system elapsed
69.47 0.61 70.64

New….

require(Matrix)
set.seed(123)
X <- Matrix(rnorm(1e6), 1000)
> print(system.time(for(i in 1:25) X%*%X))
user system elapsed
7.37 0.12 7.59
> print(system.time(for(i in 1:25) solve(X)))
user system elapsed
7.04 0.18 7.39
> print(system.time(for(i in 1:10) svd(X)))
user system elapsed
43.24 0.69 44.14

Wow!

Because my laptop is a Mac (yes, I am the happy owner of a Macbook Air), I might have a look at this site.

This directory contains versions of Rblas.dll linked against the ATLAS library:

- Athlon64_SSE3
- AthlonXP
- PII, PIII, P4, and PM (Pentium M) for various Pentiums
- C2D for Core2Duo (but not necessarily corresponding QuadCore Xeons)


The Rblas.dll in PII, PIII and AthlonXP are linked against ATLAS 3.4.1,
P4 and PM against ATLAS 3.6.0,
Athlon64_SSE3_512k against ATLAS 3.7.37, and
C2D against ATLAS 3.9.11.

Rename ...\bin\Rblas.dll and drop the new Rblas.dll into the ...\bin folder.

Using the wrong version will either not work at all or work much slower
than is possible.
Note that using a different BLAS may also result in numerical issues.