memory - C++: Fast way to read mapped file into a matrix

Question

I'm trying to read a mapped file into a matrix. The file is something like this:

name;phone;city

Luigi Rossi;02341567;Milan

Mario Bianchi;06567890;Rome

.... 

and it's quiet big. The code I've written works properly but it's not so fast:

#include <iostream>
#include <fstream>
#include <string>
#include <boost/iostreams/device/mapped_file.hpp>

using namespace std;

int main() {

    int i;
    int j=0;
    int k=0;

    vector< vector<char> > M(10000000, vector<string>(3));

    mapped_file_source file("file.csv");

    // Check if file was successfully opened
    if(file.is_open()) {

      // Get pointer to the data
      const char * c = (const char *)file.data();

      int size=file.size();

      for(i = 0; i < (size+1); i++){

       if(c[i]=='
' || i==size){
        j=j+1;
        k=0;
       }else if(c[i]==';'){
        k=k+1;
       }else{
        M[j][k]+=c[i];
       }    
     }//end for


   }//end if    

 return(0)


}

Is there a faster way? I've read something about memcyp but I don't know how to use it to speed up my code.

Answer 1

I have numerous examples doing this/similar written up on SO.

Let me list the most relevant:

• I've done quite a few of these benchmarks. Yes, for sequential freading, read/scanf have a tiny edge (see e.g. scanf/iostreams and files vs. mappings, and parsing floats, or read being slightly faster for 1-pass sequential read).

• An interesting approach is to do parsing lazily (why copy the whole input into memory? What's the point memory mapping then). The answer here shows this approach (emulating a multimap there):

  • Using boost::iostreams::mapped_file_source with std::multimap (approach #2)

In all other cases, consider slamming a Spirit Qi job on it, potentially using boost::string_ref instead of vector<char> (unless the mapped file is not "const", of course).

The string_ref is also shown int the last answer linked before. Another interesting example of this (with lazy conversions to un-escaped string values) is here How to parse mustache with Boost.Xpressive correctly?

DEMO

Here's that Qi job slammed on it:

• it parses a 994 MiB file of ~32 million lines in 2.9s into a vector of

  struct Line {
      boost::string_ref name, city;
      long id;
  };
  

• note that we parse the number, and store the strings by referring to their location in the memory map + length (string_ref)

• it pretty-prints the data from 10 random lines
• it can run as fast as 2.5s if you reserve 32m elements in the vector at once; the program does only a single memory allocation in that case.
• NOTE: on a 64 bit system, the memory representation grows larger than the input size if the average line length is less than 40 bytes. This is because a string_ref is 16 bytes.

Live On Coliru

#include <boost/fusion/adapted/struct.hpp>
#include <boost/spirit/include/qi.hpp>
#include <boost/iostreams/device/mapped_file.hpp>
#include <boost/utility/string_ref.hpp>

namespace qi = boost::spirit::qi;
using sref   = boost::string_ref;

namespace boost { namespace spirit { namespace traits {
    template <typename It>
    struct assign_to_attribute_from_iterators<sref, It, void> {
        static void call(It f, It l, sref& attr) { attr = { f, size_t(std::distance(f,l)) }; }
    };
} } }

struct Line {
    sref name, city;
    long id;
};

BOOST_FUSION_ADAPT_STRUCT(Line, (sref,name)(long,id)(sref,city))

int main() {
    boost::iostreams::mapped_file_source mmap("input.txt");

    using namespace qi;

    std::vector<Line> parsed;
    parsed.reserve(32000000);
    if (phrase_parse(mmap.begin(), mmap.end(), 
                omit[+graph] >> eol >>
                (raw[*~char_(";
")] >> ';' >> long_ >> ';' >> raw[*~char_(";
")]) % eol,
                qi::blank, parsed))
    {
        std::cout << "Parsed " << parsed.size() << " lines
";
    } else {
        std::cout << "Failed after " << parsed.size() << " lines
";
    }

    std::cout << "Printing 10 random items:
";
    for(int i=0; i<10; ++i) {
        auto& line = parsed[rand() % parsed.size()];
        std::cout << "city: '" << line.city << "', id: " << line.id << ", name: '" << line.name << "'
";
    }
}

With input generated like

do grep -v "'" /etc/dictionaries-common/words | sort -R | xargs -d\n -n 3 | while read a b c; do echo "$a $b;$RANDOM;$c"; done

The output is e.g.

Parsed 31609499 lines
Printing 10 random items:
city: 'opted', id: 14614, name: 'baronets theosophy'
city: 'denominated', id: 24260, name: 'insignia ophthalmic'
city: 'mademoiselles', id: 10791, name: 'smelter orienting'
city: 'ducked', id: 32155, name: 'encircled flippantly'
city: 'garotte', id: 3080, name: 'keeling South'
city: 'emirs', id: 14511, name: 'Aztecs vindicators'
city: 'characteristically', id: 5473, name: 'constancy Troy'
city: 'savvy', id: 3921, name: 'deafer terrifically'
city: 'misfitted', id: 14617, name: 'Eliot chambray'
city: 'faceless', id: 24481, name: 'shade forwent'