I've been using gsub
extensively lately, and I noticed that short patterns run faster than long ones, which is not surprising. Here's a fully reproducible code:
library(microbenchmark)
set.seed(12345)
n = 0
rpt = seq(20, 1461, 20)
msecFF = numeric(length(rpt))
msecFT = numeric(length(rpt))
inp = rep("aaaaaaaaaa",15000)
for (i in rpt) {
n = n + 1
print(n)
patt = paste(rep("a", rpt[n]), collapse = "")
#time = microbenchmark(func(count[1:10000,12], patt, "b"), times = 10)
timeFF = microbenchmark(gsub(patt, "b", inp, fixed=F), times = 10)
msecFF[n] = mean(timeFF$time)/1000000.
timeFT = microbenchmark(gsub(patt, "b", inp, fixed=T), times = 10)
msecFT[n] = mean(timeFT$time)/1000000.
}
library(ggplot2)
library(grid)
library(gridExtra)
axis(1,at=seq(0,1000,200),labels=T)
p1 = qplot(rpt, msecFT, xlab="pattern length, characters", ylab="time, msec",main="fixed = TRUE" )
p2 = qplot(rpt, msecFF, xlab="pattern length, characters", ylab="time, msec",main="fixed = FALSE")
grid.arrange(p1, p2, nrow = 2)
As you see, I'm looking for a pattern that contains a
replicated rpt[n]
times. The slope is positive, as expected. However, I noticed a kink at 300 characters with fixed=T
and 600 characters with fixed=F
and then the slope seems to be approximately as before (see plot below).
I suppose, it is due to memory, object size, etc. I also noticed that the longest allowed pattern
is 1463 symbols, with object size of 1552 bytes.
Can someone explain the kink better and why at 300 and 600 characters?
Added: it is worth mentioning, that most of my patterns are 5-10 characters long, which gives me on my real data (not the mock-up inp
in the example above) the following timing.
gsub, fixed = TRUE: ~50 msec per one pattern
gsub, fixed = FALSE: ~190 msec per one pattern
stringi, fixed = FALSE: ~55 msec per one pattern
gsub, fixed = FALSE, perl = TRUE: ~95 msec per one pattern
(I have 4k patterns, so total timing of my module is roughly 200 sec, which is exactly 0.05 x 4000 with gsub and fixed = TRUE. It is the fastest method for my data and patterns)
The kinks might be related to the bits required to hold patterns of that length.
There is another solution that scales much better, use the repetition operator {}
to specify how many repeats you want to find. In order to find more than 255 (8 bit integer max) you'll have to specify perl = TRUE
.
patt2 <- paste0('a{',rpt[n],'}')
timeRF <- microbenchmark(gsub(patt2, "b", inp, perl = T), times = 10)
I get speeds of around 2.1 ms per search with no penalty for pattern length. That's about 8x faster than fixed = FALSE for small pattern lengths and about 60x faster for large pattern lengths.
来源:https://stackoverflow.com/questions/27534296/gsub-speed-vs-pattern-length