R Code Optimization

1
R Code Optimization

• Matthew T. Pratola
• Dept. of Statistics and Actuarial Science
• Simon Fraser University
• 10.19.2004

2
Outline

• A quick bit about R
• Why is it slow?
• How can we speed it up?
• Conclusions

3
R Code Optimization

• R is an interpreted language
• Easy to author code to reflect an idea due to
  the friendliness of the language
• Worry about efficiency later

4
Why might it be slow?

• Computers operate on a few basic datatypes
  integers, floats, chars to name the main types
• R doesnt directly operate with these underlying
  datatypes, instead it works on abstract
  representations of these (and more complex)
  types.

5
Why might it be slow?

• Why? So strange things like NAs, data frames,
  etc, can be easily created, accessed and
  manipulated by the end user

6
Why might it be slow?

• Read abstract representations gt think more
  memory used to represent a given unit of data
• This means that for every operation you do on
  your data, your computer must spend extra time
  moving around all this memory, much of which
  doesnt directly have anything to do with the
  value of interest

7
How to make it faster?

• Standard approaches
• Investigate algorithmic design of your program
• Optimize memory access patterns
• Others

8
How?

• Avoid using dataframes within your code
• Simple example
• mmatrix(nrow2000,ncol1)
• for(i in (12000))mi,11
• vs.
• mas.data.frame(m)
• for(i in (12000))mi,11

9
How?

• Difference for the above was 0.05s vs. 5.47s
• Fix Only convert your data to a dataframe at
  the end of the program ie for the user

10
How?

• Another good trick avoid the use of rbind/cbind
  if possible during long (intensive) loops
• Instead, pre-allocate the entire (or, if not
  possible, a very large) array initialized to
  NAs, then directly assign the values using a row
  (or col) counter

11
How?

• Eg
• xc(1,1,1,1)
• for(i in 15000)xrbind(x,c(i,i,i,i))
• vs.
• xmatrix(nrow5000,ncol4)
• for(i in 15000)xi,c(i,i,i,i)
• 0.15s vs. 12.45s

12
How?

• Another trick often it may be the case that you
  have a main loop in your code, perhaps updating
  many matrices
• But, it could be that there is no interdependence
  amongst the matrices you are updating

13
How?

• Consider
• for()
• blah
• blah2
• vs.
• for() blah
• for() blah2

14
How?

• This second approach can often yield a reasonable
  gain in a very long, intensive loop
• Real compilers (C, Fortran, etc) do this
  automatically, but R does not

15
How?

• A final idea avoid excessive use of string
  manipulation functions, such as
• as.numeric(unlist(strsplit(myvar,split)))
• Working around excessive use of this type of code
  can provide some improvement

16
How?

• Other more detailed optimizations can yield some
  improvement, although past these simple
  approaches, the percentage improvement will
  quickly fall to single digit levels
• Recognize that many basic R operations are
  optimally coded in C, so if you write your code
  to fit well with these operations, you will
  retain execution speed

17
How? system.time

• Quick and easy way to test, say, one particular
  function
• gc() forces a garbage collection event
• system.time(myfunction)
• gives you output like
• 1 0.64 0.00 0.66 NA NA
• Format (user)(system)(elapsed)(NA)(NA)

18
How? Profiling

• Use Rprof to gain insight into where your code is
  slow
• Rprof(Rprof.dat)
• ltcodegt
• Rprof(NULL)
• Get profile output by running
• R CMD Rprof Rprof.dat gt profile.txt

19
How? - Profiling

• You should always use a high-level function
  when you are profiling, or else Rprof may not
  report the time you expect it will report
• Not sure if this is available on windows

20
Conclusion

• Speedup is possible in R using a few tricks
• Sometimes these are easy to implement, sometimes
  they are not
• Try to achieve an idea between code that conveys
  an idea and code that runs quickly
• Try simple approaches that are within your
  comfort zone

21
Questions?