The previous post looked at using the Vectorize() function to, well, vectorize, our Rcpp IPv4 functions. While this is a completely acceptable practice, we can perform the vectorization 100% in Rcpp/C++. We’ve included both the original Rcpp IPv4 functions and the new Rcpp-vectorized functions together to show the minimal differences between them:

#include <Rcpp.h> 
#include <boost/asio/ip/address_v4.hpp>

using namespace Rcpp; 
using namespace boost::asio::ip;

// Rcpp/C++ vectorized routines

// [[Rcpp::export]]
NumericVector rcpp_rinet_pton (CharacterVector ip) {

  int ipCt = ip.size(); // how many elements in vector

  NumericVector ipInt(ipCt); // allocate new numeric vector

  // CONVERT ALL THE THINGS!
  for (int i=0; i<ipCt; i++) {
    ipInt[i] = address_v4::from_string(ip[i]).to_ulong();
  }

  return(ipInt);
}

// [[Rcpp::export]]
CharacterVector rcpp_rinet_ntop (NumericVector ip) {

  int ipCt = ip.size();

  CharacterVector ipStr(ipCt); // allocate new character vector
  // CONVERT ALL THE THINGS!
  for (int i=0; i<ipCt; i++) {
    ipStr[i] = address_v4(ip[i]).to_string();
  }

  return(ipStr);

}

// orignial single-element vector routines we'll vectorize with Vectorize()

// [[Rcpp::export]]
unsigned long rinet_pton (CharacterVector ip) { 
  return(boost::asio::ip::address_v4::from_string(ip[0]).to_ulong());
}

// [[Rcpp::export]]
CharacterVector rinet_ntop (unsigned long addr) {
  return(boost::asio::ip::address_v4(addr).to_string());
}

We’ve merely wrapped a for loop around the original code and built the result vectors in Rcpp, relying on the object-oriented nature of C++ for proper value conversion+assignment. The pure-R+Vectorize()‘d code (from the examples in the book) is below, since we’re going to pit all three in a head-to-head performance competition.

# Vectorize() the single-element vector routines
v_rinet_pton <- Vectorize(rinet_pton, USE.NAMES=FALSE)
v_rinet_ntop <- Vectorize(rinet_ntop, USE.NAMES=FALSE)

# pure R version with Vectorize()
ip2long <- Vectorize(function(ip) {
  ips <- unlist(strsplit(ip, '.', fixed=TRUE))
  octet <- function(x,y) bitOr(bitShiftL(x, 8), y)
  Reduce(octet, as.integer(ips))
}, USE.NAMES=FALSE)

long2ip <- Vectorize(function(longip) {
  octet <- function(nbits) bitAnd(bitShiftR(longip, nbits), 0xFF)
  paste(Map(octet, c(24,16,8,0)), sep="", collapse=".")
}, USE.NAMES=FALSE)

Now, we’ll read in a file of ~8,000 IPv4 addresses, make them into integers and then use the microbenchmark package to profile the to/from conversion of all three versions of the routines.

# read in ~8K IP address strings & make ints for our benchmark
ips <- read.table("data/ips.dat", header=FALSE, stringsAsFactors=FALSE)
ints <- rcpp_rinet_pton(ips$V1)

# run a benchmark 100 times per routine, giving plenty of "ramp up" time
mb <- microbenchmark(rcpp_ints <- rcpp_rinet_pton(ips$V1), 
                     rcpp_chars <- rcpp_rinet_ntop(ints),
                     v_ints <- v_rinet_pton(ips$V1),
                     v_chars <- v_rinet_ntop(ints), 
                     r_ints <- ip2long(ips$V1),
                     r_chars <- long2ip(ints),
                     control=list(warmup=20),
                     times=100, unit="s")

Then, we’ll take a look at the results (all times are in seconds):

If we just look at the median values, we can see that the conversion to integer takes:

and, the conversion to character takes:

But, a visualization is (often) worth a dozen tables, so we’ll take the test results and make a violin plot (which is just a more granular boxplot). Note that the plot is on a log scale, so the differences between each set of comparisons are actually much larger than your eye will initially comprehend (hence the inclusion of the above tables).

It’s often difficult for us to grok fractional seconds, so let’s do some basic math to see how long each method would take to process 1 billion IP addresses. We’ll use the median values from above and compare the results in a simple bar chart:

The fully vectorized Rcpp versions are the clear “winners” and will let us scale our IPv4 address conversions to millions, billions or trillions of operations without having to rely on other scripting languages. We can use this base as foundation for a complete IP address S4 class that we’ll cover in future posts.

You can find the Rmd source that helped generate this post over at github along with the data file.