使用 Boost::spirit 编写的解析器的性能问题

Question

我想解析一个看起来像这样的文件（FASTA-like 文本格式）：

    >InfoHeader
    "Some text sequence that has a line break after every 80 characters"
    >InfoHeader
    "Some text sequence that has a line break after every 80 characters"
    ...

例如：

    >gi|31563518|ref|NP_852610.1| microtubule-associated proteins 1A/1B light chain 3A isoform b [Homo sapiens]
    MKMRFFSSPCGKAAVDPADRCKEVQQIRDQHPSKIPVIIERYKGEKQLPVLDKTKFLVPDHVNMSELVKI
    IRRRLQLNPTQAFFLLVNQHSMVSVSTPIADIYEQEKDEDGFLYMVYASQETFGFIRENE

我为此使用 boost::spirit 编写了一个解析器。解析器正确地将标题行和以下文本序列存储在 std::vector< std::pair< string, string >> 中，但对于较大的文件需要很长时间（100MB 文件需要 17 秒）。作为比较，我编写了一个没有 boost::spirit 的程序（只是 STL 函数），它只是将 100MB 文件的每一行复制到 std::vector 中。整个过程不到一秒钟。用于比较的“程序”没有达到目的，但我认为解析器不应该花那么多时间......

我知道周围还有很多其他 FASTA 解析器，但我很好奇为什么我的代码很慢。

.hpp 文件：

#include <boost/filesystem/path.hpp>

namespace fs = boost::filesystem;


class FastaReader {

public:
    typedef std::vector< std::pair<std::string, std::string> > fastaVector;

private:
    fastaVector fV;
    fs::path file;  

public:
    FastaReader(const fs::path & f);
    ~FastaReader();

    const fs::path & getFile() const;
    const fastaVector::const_iterator getBeginIterator() const;
    const fastaVector::const_iterator getEndIterator() const;   

private:
    void parse();

};

还有.cpp文件：

#include <iomanip>
#include <boost/date_time/posix_time/posix_time.hpp>
#include <boost/filesystem/fstream.hpp>
#include <boost/filesystem/operations.hpp>
#include <boost/filesystem/path.hpp>
#include <boost/spirit/include/classic_position_iterator.hpp>
#include <boost/spirit/include/phoenix_bind.hpp>
#include <boost/spirit/include/phoenix_core.hpp>
#include <boost/spirit/include/phoenix_fusion.hpp>
#include <boost/spirit/include/phoenix_operator.hpp>
#include <boost/spirit/include/qi.hpp>
#include "fastaReader.hpp"


using namespace std;

namespace fs = boost::filesystem;
namespace qi = boost::spirit::qi;
namespace pt = boost::posix_time;

template <typename Iterator, typename Skipper>
struct FastaGrammar : qi::grammar<Iterator, FastaReader::fastaVector(), qi::locals<string>, Skipper> {
    qi::rule<Iterator> infoLineStart;
    qi::rule<Iterator> inputEnd;
    qi::rule<Iterator> lineEnd;
    qi::rule<Iterator, string(), Skipper> infoLine;
    qi::rule<Iterator, string(), Skipper> seqLine;
    qi::rule<Iterator, FastaReader::fastaVector(), qi::locals<string>, Skipper> fasta;


    FastaGrammar() : FastaGrammar::base_type(fasta, "fasta") {
        using boost::spirit::standard::char_;
        using boost::phoenix::bind;
        using qi::eoi;
        using qi::eol;
        using qi::lexeme;
        using qi::_1;
        using qi::_val;
        using namespace qi::labels;

        infoLineStart = char_('>');
        inputEnd = eoi;

        /* grammar */       
        infoLine = lexeme[*(char_ - eol)];
        seqLine = *(char_ - infoLineStart);

        fasta = *(infoLineStart > infoLine[_a = _1] 
            > seqLine[bind(&FastaGrammar::addValue, _val, _a, _1)]
            )
            > inputEnd
        ;

        infoLineStart.name(">");
        infoLine.name("sequence identifier");
        seqLine.name("sequence");

    }

    static void addValue(FastaReader::fastaVector & fa, const string & info, const string & seq) {
        fa.push_back(make_pair(info, seq));
    }
};


FastaReader::FastaReader(const fs::path & f) {
    this->file = f; 
    this->parse();
}


FastaReader::~FastaReader() {}


const fs::path & FastaReader::getFile() const {
    return this->file;
}


const FastaReader::fastaVector::const_iterator FastaReader::getBeginIterator() const {
    return this->fV.cbegin();
}


const FastaReader::fastaVector::const_iterator FastaReader::getEndIterator() const {
    return this->fV.cend();
}


void FastaReader::parse() {
    if ( this->file.empty() ) throw string("FastaReader: No file specified.");
    if ( ! fs::is_regular_file(this->file) ) throw (string("FastaReader: File not found: ") + this->file.string());

    typedef boost::spirit::istream_iterator iterator_type;
    typedef boost::spirit::classic::position_iterator2<iterator_type> pos_iterator_type;
    typedef FastaGrammar<pos_iterator_type, boost::spirit::ascii::space_type> fastaGr;

    fs::ifstream fin(this->file);
    if ( ! fin.is_open() ) {
        throw (string("FastaReader: Access denied: ") + this->file.string());
    }

    fin.unsetf(ios::skipws);

    iterator_type begin(fin);
    iterator_type end;

    pos_iterator_type pos_begin(begin, end, this->file.string());
    pos_iterator_type pos_end;

    fastaGr fG;
    try {
        std::cerr << "Measuring: Parsing." << std::endl;
        const pt::ptime startMeasurement = pt::microsec_clock::universal_time();

        qi::phrase_parse(pos_begin, pos_end, fG, boost::spirit::ascii::space, this->fV);

        const pt::ptime endMeasurement = pt::microsec_clock::universal_time();
        pt::time_duration duration (endMeasurement - startMeasurement);
        std::cerr << duration <<  std::endl;
    } catch (std::string str) {
        cerr << "error message: " << str << endl;
    }   
}

所以语法做了以下事情： 它查找“>”符号，然后存储所有后续字符，直到检测到 EOL。在 EOL 之后，文本序列开始并在检测到“>”符号时结束。然后通过调用 FastaReader::addValue() 将两个字符串（标题行和文本序列）存储在 std::vector 中。

我使用带有 -O2 和 -std=c++11 标志的 g++ 版本 4.8.2 编译了我的程序。

那么我的代码中的性能问题在哪里？

Answer 1

_{上一个：Step 3: MOAR FASTER WITH ZERO-COPY
返回Step 1. Cleaning up + Profiling}

第 4 步：删除位置迭代器

由于您不使用它，我们可以删除有状态迭代器，这可能会抑制很多优化（并且在 the profiler output 中间接可见）

Live On Coliru

#define BOOST_SPIRIT_USE_PHOENIX_V3
#include <boost/filesystem/path.hpp>
#include <boost/utility/string_ref.hpp>
#include <boost/iostreams/device/mapped_file.hpp>
namespace io = boost::iostreams;
namespace fs = boost::filesystem;


class FastaReader {

public:
    typedef std::pair<boost::string_ref, boost::string_ref> Entry;
    typedef std::vector<Entry> Data;

private:
    Data fV;
    fs::path file;  

public:
    FastaReader(const fs::path & f);
    ~FastaReader();

    const fs::path & getFile() const;
    const Data::const_iterator begin() const;
    const Data::const_iterator end() const;   

private:
    io::mapped_file_source mmap;
    void parse();

};

#include <iomanip>
#include <boost/date_time/posix_time/posix_time.hpp>
#include <boost/filesystem/fstream.hpp>
#include <boost/filesystem/operations.hpp>
#include <boost/filesystem/path.hpp>

#include <boost/spirit/include/qi.hpp>
#include <boost/spirit/include/phoenix.hpp>
#include <boost/fusion/adapted/std_pair.hpp>
//#include "fastaReader.hpp"

#include <boost/iostreams/device/mapped_file.hpp>

using namespace std;

namespace fs = boost::filesystem;
namespace qi = boost::spirit::qi;
namespace pt = boost::posix_time;
namespace io = boost::iostreams;

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

template <typename Iterator>
struct FastaGrammar : qi::grammar<Iterator, FastaReader::Data()> {

    FastaGrammar() : FastaGrammar::base_type(fasta) {
        using namespace qi;
        using boost::phoenix::construct;
        using boost::phoenix::begin;
        using boost::phoenix::size;

        entry = ('>' >> raw[ *~char_('\n') ] >> '\n' >> raw[ *~char_('>') ]);
        fasta = *entry >> *eol >> eoi ;

        BOOST_SPIRIT_DEBUG_NODES((fasta)(entry));
    }
  private:
    qi::rule<Iterator, FastaReader::Data()>  fasta;
    qi::rule<Iterator, FastaReader::Entry()> entry;
};

FastaReader::FastaReader(const fs::path & f) : file(f), mmap(file.c_str()) {
    parse();
}

FastaReader::~FastaReader() {}

const fs::path & FastaReader::getFile() const {
    return this->file;
}


const FastaReader::Data::const_iterator FastaReader::begin() const {
    return this->fV.cbegin();
}


const FastaReader::Data::const_iterator FastaReader::end() const {
    return this->fV.cend();
}

void FastaReader::parse() {
    if (this->file.empty())                throw std::runtime_error("FastaReader: No file specified.");
    if (! fs::is_regular_file(this->file)) throw std::runtime_error(string("FastaReader: File not found: ") + this->file.string());

    typedef char const*                  iterator_type;
    typedef FastaGrammar<iterator_type>  fastaGr;

    static const fastaGr fG{};
    try {
        std::cerr << "Measuring: Parsing." << std::endl;
        const pt::ptime startMeasurement = pt::microsec_clock::universal_time();

        iterator_type first(mmap.data()), last(mmap.end());
        qi::phrase_parse(first, last, fG, boost::spirit::ascii::space, this->fV);

        const pt::ptime endMeasurement = pt::microsec_clock::universal_time();
        pt::time_duration duration (endMeasurement - startMeasurement);
        std::cerr << duration <<  std::endl;
    } catch (std::exception const& e) {
        cerr << "error message: " << e.what() << endl;
    }   
}

int main() {
    FastaReader reader("input.txt");

    for (auto& e : reader) std::cout << '>' << e.first << '\n' << e.second << "\n\n";
}

现在速度提高了 74.8 倍。

$ time ./test | head -n4
Measuring: Parsing.
00:00:00.194432

Answer 2

_{下一个：Step 2. Faster with mmap}

步骤 1. 清理 + 分析

你应该避免他们引入类型擦除的许多规则。

如果你的输入是理智的，你可以不用船长（无论如何，行尾很重要，所以跳过它们是没有意义的）。

使用融合适应而不是帮助器来构建新的对：

这还不是最佳的，但更干净：

$ ./test1
Measuring: Parsing.
00:00:22.681605

通过减少移动部件和间接来稍微提高效率：

Live On Coliru

#include <boost/filesystem/path.hpp>

namespace fs = boost::filesystem;

class FastaReader {    
public:
    typedef std::pair<std::string, std::string> Entry;
    typedef std::vector<Entry> Data;

private:
    Data fV;
    fs::path file;  

public:
    FastaReader(const fs::path & f);
    ~FastaReader();

    const fs::path & getFile() const;
    const Data::const_iterator begin() const;
    const Data::const_iterator end() const;   

private:
    void parse();    
};

#include <iomanip>
#include <boost/date_time/posix_time/posix_time.hpp>
#include <boost/filesystem/fstream.hpp>
#include <boost/filesystem/operations.hpp>
#include <boost/filesystem/path.hpp>

#include <boost/spirit/include/classic_position_iterator.hpp>
#include <boost/spirit/include/qi.hpp>
#include <boost/fusion/adapted/std_pair.hpp>
//#include "fastaReader.hpp"

using namespace std;

namespace fs = boost::filesystem;
namespace qi = boost::spirit::qi;
namespace pt = boost::posix_time;

template <typename Iterator>
struct FastaGrammar : qi::grammar<Iterator, FastaReader::Data()> {
    qi::rule<Iterator, FastaReader::Data()> fasta;

    FastaGrammar() : FastaGrammar::base_type(fasta) {
        using namespace qi;

        fasta = *('>' >> *~char_('\n') >> '\n' 
                      >> *~char_('>')) 
                >> *eol
                >> eoi
                ;

        BOOST_SPIRIT_DEBUG_NODES((fasta));
    }
};


FastaReader::FastaReader(const fs::path & f) : file(f) {
    parse();
}

FastaReader::~FastaReader() {}

const fs::path & FastaReader::getFile() const {
    return this->file;
}

const FastaReader::Data::const_iterator FastaReader::begin() const {
    return this->fV.cbegin();
}

const FastaReader::Data::const_iterator FastaReader::end() const {
    return this->fV.cend();
}

void FastaReader::parse() {
    if (this->file.empty())                throw std::runtime_error("FastaReader: No file specified.");
    if (! fs::is_regular_file(this->file)) throw std::runtime_error(string("FastaReader: File not found: ") + this->file.string());

    typedef boost::spirit::istream_iterator                           iterator_type;
    typedef boost::spirit::classic::position_iterator2<iterator_type> pos_iterator_type;
    typedef FastaGrammar<pos_iterator_type>                           fastaGr;

    fs::ifstream fin(this->file);
    if (!fin) {
        throw std::runtime_error(string("FastaReader: Access denied: ") + this->file.string());
    }

    static const fastaGr fG{};
    try {
        std::cerr << "Measuring: Parsing." << std::endl;
        const pt::ptime startMeasurement = pt::microsec_clock::universal_time();

        pos_iterator_type first(iterator_type{fin >> std::noskipws}, {}, file.string());
        qi::phrase_parse<pos_iterator_type>(first, {}, fG, boost::spirit::ascii::space, this->fV);

        const pt::ptime endMeasurement = pt::microsec_clock::universal_time();
        pt::time_duration duration (endMeasurement - startMeasurement);
        std::cerr << duration <<  std::endl;
    } catch (std::exception const& e) {
        cerr << "error message: " << e.what() << endl;
    }   
}

int main() {
    std::ios::sync_with_stdio(false);

    FastaReader reader("input.txt");

    //for (auto& e : reader) std::cout << '>' << e.first << '\n' << e.second << "\n\n";
}

这仍然很慢。让我们看看需要这么长时间：

这很好，但几乎没有告诉我们需要知道什么。然而，这确实是：前 N 次消费者是

所以大部分时间都花在 istream 迭代和多通道适配器上。您可能会争辩说，可以通过不时刷新一次（每行？）来优化多通道适配器，但实际上，我们不希望将其绑定到（流）缓冲区上的整个流和运算符。

所以，我虽然让我们使用映射文件：

下一个：Step 2. Faster with mmap

Answer 3

_{上一个：Step 2. Faster with mmap
下一个：Step 4: Dropping the position iterator}

第 3 步：使用零拷贝加快速度

让我们避免分配！如果我们将文件映射移到 FastaReader 类中，我们可以直接指向映射中的数据，而不是一直复制字符串。

使用 boost::string_ref 例如在这里描述：C++: Fast way to read mapped file into a matrix你可以做

Live On Coliru

#define BOOST_SPIRIT_USE_PHOENIX_V3
#include <boost/filesystem/path.hpp>
#include <boost/utility/string_ref.hpp>
#include <boost/iostreams/device/mapped_file.hpp>
namespace io = boost::iostreams;
namespace fs = boost::filesystem;

class FastaReader {

public:
    typedef std::pair<boost::string_ref, boost::string_ref> Entry;
    typedef std::vector<Entry> Data;

private:
    Data fV;
    fs::path file;  

public:
    FastaReader(const fs::path & f);
    ~FastaReader();

    const fs::path & getFile() const;
    const Data::const_iterator begin() const;
    const Data::const_iterator end() const;   

private:
    io::mapped_file_source mmap;
    void parse();

};

#include <iomanip>
#include <boost/date_time/posix_time/posix_time.hpp>
#include <boost/filesystem/fstream.hpp>
#include <boost/filesystem/operations.hpp>
#include <boost/filesystem/path.hpp>

#include <boost/spirit/include/classic_position_iterator.hpp>
#include <boost/spirit/include/qi.hpp>
#include <boost/spirit/include/phoenix.hpp>
#include <boost/fusion/adapted/std_pair.hpp>
//#include "fastaReader.hpp"

#include <boost/iostreams/device/mapped_file.hpp>

using namespace std;

namespace fs = boost::filesystem;
namespace qi = boost::spirit::qi;
namespace pt = boost::posix_time;
namespace io = boost::iostreams;

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

template <typename Iterator>
struct FastaGrammar : qi::grammar<Iterator, FastaReader::Data()> {

    FastaGrammar() : FastaGrammar::base_type(fasta) {
        using namespace qi;
        using boost::phoenix::construct;
        using boost::phoenix::begin;
        using boost::phoenix::size;

        entry = ('>' >> raw[ *~char_('\n') ] >> '\n' >> raw[ *~char_('>') ]);
        fasta = *entry >> *eol >> eoi ;

        BOOST_SPIRIT_DEBUG_NODES((fasta)(entry));
    }
  private:
    qi::rule<Iterator, FastaReader::Data()>  fasta;
    qi::rule<Iterator, FastaReader::Entry()> entry;
};

FastaReader::FastaReader(const fs::path & f) : file(f), mmap(file.c_str()) {
    parse();
}

FastaReader::~FastaReader() {}

const fs::path & FastaReader::getFile() const {
    return this->file;
}


const FastaReader::Data::const_iterator FastaReader::begin() const {
    return this->fV.cbegin();
}


const FastaReader::Data::const_iterator FastaReader::end() const {
    return this->fV.cend();
}

void FastaReader::parse() {
    if (this->file.empty())                throw std::runtime_error("FastaReader: No file specified.");
    if (! fs::is_regular_file(this->file)) throw std::runtime_error(string("FastaReader: File not found: ") + this->file.string());

    typedef char const*                                               iterator_type;
    typedef boost::spirit::classic::position_iterator2<iterator_type> pos_iterator_type;
    typedef FastaGrammar<pos_iterator_type>                           fastaGr;

    static const fastaGr fG{};
    try {
        std::cerr << "Measuring: Parsing." << std::endl;
        const pt::ptime startMeasurement = pt::microsec_clock::universal_time();

        pos_iterator_type first(iterator_type{mmap.data()}, iterator_type{mmap.end()}, file.string());
        qi::phrase_parse<pos_iterator_type>(first, {}, fG, boost::spirit::ascii::space, this->fV);

        const pt::ptime endMeasurement = pt::microsec_clock::universal_time();
        pt::time_duration duration (endMeasurement - startMeasurement);
        std::cerr << duration <<  std::endl;
    } catch (std::exception const& e) {
        cerr << "error message: " << e.what() << endl;
    }   
}

int main() {
    FastaReader reader("input.txt");

    for (auto& e : reader) std::cout << '>' << e.first << '\n' << e.second << "\n\n";
}

这确实快了 4.8 倍：

$ ./test3 | head -n4
Measuring: Parsing.
00:00:04.577123
>gi|31563518|ref|NP_852610.1| microtubule-associated proteins 1A/1B light chain 3A isoform b [Homo sapiens]
MKMRFFSSPCGKAAVDPADRCKEVQQIRDQHPSKIPVIIERYKGEKQLPVLDKTKFLVPDHVNMSELVKI
IRRRLQLNPTQAFFLLVNQHSMVSVSTPIADIYEQEKDEDGFLYMVYASQETFGFIRENE

下一个：Step 4: Dropping the position iterator

Answer 4

_{上一个：Step 1. Cleaning up + Profiling
下一个：Step 3: MOAR FASTER WITH ZERO-COPY}

第 2 步。使用mmap 更快

Live On Coliru

#include <boost/filesystem/path.hpp>

namespace fs = boost::filesystem;


class FastaReader {

public:
    typedef std::pair<std::string, std::string> Entry;
    typedef std::vector<Entry> Data;

private:
    Data fV;
    fs::path file;  

public:
    FastaReader(const fs::path & f);
    ~FastaReader();

    const fs::path & getFile() const;
    const Data::const_iterator begin() const;
    const Data::const_iterator end() const;   

private:
    void parse();

};

#include <iomanip>
#include <boost/date_time/posix_time/posix_time.hpp>
#include <boost/filesystem/fstream.hpp>
#include <boost/filesystem/operations.hpp>
#include <boost/filesystem/path.hpp>

#include <boost/spirit/include/classic_position_iterator.hpp>
#include <boost/spirit/include/qi.hpp>
#include <boost/fusion/adapted/std_pair.hpp>
//#include "fastaReader.hpp"

#include <boost/iostreams/device/mapped_file.hpp>

using namespace std;

namespace fs = boost::filesystem;
namespace qi = boost::spirit::qi;
namespace pt = boost::posix_time;
namespace io = boost::iostreams;

template <typename Iterator>
struct FastaGrammar : qi::grammar<Iterator, FastaReader::Data()> {
    qi::rule<Iterator, FastaReader::Data()> fasta;

    FastaGrammar() : FastaGrammar::base_type(fasta) {
        using namespace qi;

        fasta = *('>' >> *~char_('\n') >> '\n' 
                      >> *~char_('>')) 
                >> *eol
                >> eoi
                ;

        BOOST_SPIRIT_DEBUG_NODES((fasta));
    }
};


FastaReader::FastaReader(const fs::path & f) : file(f) {
    parse();
}

FastaReader::~FastaReader() {}

const fs::path & FastaReader::getFile() const {
    return this->file;
}


const FastaReader::Data::const_iterator FastaReader::begin() const {
    return this->fV.cbegin();
}


const FastaReader::Data::const_iterator FastaReader::end() const {
    return this->fV.cend();
}

void FastaReader::parse() {
    if (this->file.empty())                throw std::runtime_error("FastaReader: No file specified.");
    if (! fs::is_regular_file(this->file)) throw std::runtime_error(string("FastaReader: File not found: ") + this->file.string());

    typedef char const*                                               iterator_type;
    typedef boost::spirit::classic::position_iterator2<iterator_type> pos_iterator_type;
    typedef FastaGrammar<pos_iterator_type>                           fastaGr;

    io::mapped_file_source mmap(file.c_str());

    static const fastaGr fG{};
    try {
        std::cerr << "Measuring: Parsing." << std::endl;
        const pt::ptime startMeasurement = pt::microsec_clock::universal_time();

        pos_iterator_type first(iterator_type{mmap.data()}, iterator_type{mmap.end()}, file.string());
        qi::phrase_parse<pos_iterator_type>(first, {}, fG, boost::spirit::ascii::space, this->fV);

        const pt::ptime endMeasurement = pt::microsec_clock::universal_time();
        pt::time_duration duration (endMeasurement - startMeasurement);
        std::cerr << duration <<  std::endl;
    } catch (std::exception const& e) {
        cerr << "error message: " << e.what() << endl;
    }   
}

int main() {
    FastaReader reader("input.txt");

    //for (auto& e : reader) std::cout << '>' << e.first << '\n' << e.second << "\n\n";
}

确实在我的系统上它大约 快 3 倍（输入为 229 MiB）：

$ ./mapped_file_source
Measuring: Parsing.
00:00:07.385787

下一个：Step 3: MOAR FASTER WITH ZERO-COPY