[R] Speed up code with for() loop

[Jeremy Hetzel]

Hans,

You could parallelize it with the multicore package.  The only other thing I 
can think of is to use calls to .Internal().  But be vigilant, as this might 
not be good advice.  ?.Internal warns that only true R wizards should even 
consider using the function.  First, an example with .Internal() calls, 
later mutlicore.  For me, the following reduces elapsed time by about 9% on 
Windows 7 and by about 20% on today's new Ubuntu Natty.

## Set number of replicates
n <- 10000

## Your example
set.seed(1)
time.one <- Sys.time()
Error<-rnorm(n, mean=0, sd=0.05)
estimate<-(log(1.1)-Error)
DCF_korrigiert<-(1/(exp(1/(exp(0.5*(-estimate)^2/(0.05^2))*sqrt(2*pi/(0.05^2))*(1-pnorm(0,((-estimate)/(0.05^2)),sqrt(1/(0.05^2))))))-1))
D<-n
Delta_ln<-rep(0,D)
for(i in 1:D)
Delta_ln[i]<-(log(mean(sample(DCF_korrigiert,D,replace=TRUE))/(1/0.10)))
time.one <- Sys.time() - time.one

## A few modifications with .Internal()
set.seed(1)
time.two <- Sys.time()
Error <- rnorm(n, mean = 0, sd = 0.05)
estimate <- (log(1.1) - Error)
DCF_korrigiert <- (1 / (exp(1 / (exp(0.5 * (-estimate)^2 / (0.05^2)) * sqrt(
2* pi / (0.05^2)) * (1 - pnorm(0,((-estimate) / (0.05^2)), sqrt(1 / 
(0.05^2))))))-1))
D <- n
Delta_ln2 <- numeric(length = D)
Delta_ln2 <- vapply(Delta_ln2, function(x)
{
log(.Internal(mean(DCF_korrigiert[.Internal(
sample(D, D, replace = T, prob = NULL))])) / (1 / 0.10))
}, FUN.VALUE = 1)
time.two <- Sys.time() - time.two


## Compare
all.equal(Delta_ln, Delta_ln2)
time.one
time.two
as.numeric(time.two) / as.numeric(time.one)






Then you could parallelize it with multicore's parallel() function:

## Try multicore
require(multicore)
set.seed(1)
time.three <- Sys.time()
Error <- rnorm(n, mean = 0, sd = 0.05)
estimate <- (log(1.1) - Error)
DCF_korrigiert <- (1 / (exp(1 / (exp(0.5 * (-estimate)^2 / (0.05^2)) * sqrt(
2* pi / (0.05^2)) * (1 - pnorm(0,((-estimate) / (0.05^2)), sqrt(1 / 
(0.05^2))))))-1))
D <- n/2
Delta_ln3 <- numeric(length = D)
Delta_ln3.1 <- parallel(vapply(Delta_ln3, function(x)
{
log(.Internal(mean(DCF_korrigiert[.Internal(
sample(D, D, replace = T, prob = NULL))])) / (1 / 0.10))
}, FUN.VALUE = 1), mc.set.seed = T)
Delta_ln3.2 <- parallel(vapply(Delta_ln3, function(x)
{
log(.Internal(mean(DCF_korrigiert[.Internal(
sample(D, D, replace = T, prob = NULL))])) / (1 / 0.10))
}, FUN.VALUE = 1), mc.set.seed = T)
results <- collect(list(Delta_ln3.1, Delta_ln3.2))
names(results) <- NULL
Delta_ln3 <- do.call("append", results)
time.three <- Sys.time() - time.three


## Compare
# Results won't be equal due to the different way 
# parallel() handles set.seed() randomization
all.equal(Delta_ln, Delta_ln3)
time.one
time.two
time.three
as.numeric(time.three) / as.numeric(time.one)



Combining parallel() with the .Internal calls reduces the elapsed time by 
about 70% on Ubuntu Natty.  Multicore is not available for Windows, or at 
least not easily available for Windows.  

But maybe the true R wizards have better ideas.


Jeremy