MPI 和提升 multiprecision/gmp

MPI and boost multiprecision/gmp

我厌倦了将 DBL 类型数组转换为 char 数组,使用 MPI_Bcast 广播它并转换回 DBL 数组。 DBL 数组可以是以下任何一个:

  1. 双倍
  2. 长双
  3. mpf_float_50(就是boost/multiprecision/包中定义的类型)。

前两种类型工作正常:

mpicxx -D_DBL main.cpp -o main && mpirun -np 2 main

mpicxx -D_LDBL main.cpp -o main && mpirun -np 2 main

给予

before for rank=0
0
0.1
after for rank=1
0
0.1
after for rank=0
0
0.1

但是

mpicxx -D_EDBL50 main.cpp -o main -lgmp && mpirun -np 2 main
one has
before for rank=0
0
0.1
after for rank=0
0
0.1
after for rank=1
0

[Fominskoe:06235] *** Process received signal ***
[Fominskoe:06235] Signal: Segmentation fault (11)
[Fominskoe:06235] Signal code: Address not mapped (1)
[Fominskoe:06235] Failing at address: 0x55892d02d450
[Fominskoe:06235] [ 0] /lib/x86_64-linux-gnu/libc.so.6(+0x3ef20)
[0x7f26e6333f20]
[Fominskoe:06235] [ 1] /usr/lib/x86_64-linux-
gnu/libgmp.so.10(__gmpn_copyi+0x4d)[0x7f26e71f5213]
[Fominskoe:06235]

* 错误消息结束 *


mpirun 注意到节点 Fominskoe 上 PID 为 0 的进程等级 1 在信号 11(分段错误)上退出。

我注意到最后的选择 mpirun -np 1 main

给出正确答案:

before for rank=0
0
0.1
after for rank=0
0
0.1
    #include "mpi.h"

    #if defined(_DBL)
    typedef double DBL;
    #endif    

    #if defined(_LDBL)
    typedef long double DBL;
    #endif 

    #if defined(_EDBL50)
    #include <boost/multiprecision/gmp.hpp>
    using namespace boost::multiprecision;
    typedef mpf_float_50 DBL;
    #endif 

    using namespace std;

    int main(int argc, char* argv[])
    {

    MPI::Init (argc, argv);
    int proc_num = MPI::COMM_WORLD.Get_size ( );
    int my_rank  = MPI::COMM_WORLD.Get_rank ( );

    int N=2;

    DBL  DB[N];
    int  CN=N*sizeof(DBL);
    char CH[CN];

    if ( !my_rank ){

        cout<<"before for rank="<<my_rank<<endl;
        for (int i=0; i<N; i++) {               // init array
            DB[i]=i*0.1;
            cout<<DB[i]<<endl;
        }

        char* ptr=(char*)(&DB[0]);

        for (int i=0; i<CN; i++) 
            CH[i]=*ptr++;

        for (int i=0; i<N; i++) // clean
            DB[i]=0;

    }

    MPI_Bcast (CH, CN, MPI_CHAR, 0, MPI_COMM_WORLD);

    int ii=0;
    DBL* V;

    cout<<"\nafter for rank="<<my_rank<<endl;
    for (int i=0; i<N; i++) {
        V=(DBL*)(&CH[ii]); 
        DB[i]=*V;
        cout<< DB[i]<<endl;
        ii+=sizeof(DBL);
        //        if (my_rank)
        //        break;
    }

    MPI::Finalize();

    return 0;
}

如果移除 MPI,转换 DBL->char->DBL 适用于所有三种类型。

Ubuntu 18.04, gcc, mpicxx -- 打开 MPI C++ 包装器编译器、libboost-all-dev、libboost-tools-dev

有什么想法吗?

正如评论所说,你只能按位序列化trivial types

    char *ptr = (char *)(&DB[0]);

    for (int i = 0; i < CN; i++)
        CH[i] = *ptr++;

这不是转换。这是一个实现按位复制的重新解释转换,即 Undefined Behaviour 用于非平凡的多精度类型。

你很幸运,因为你想使用 MPI 和 Boost,你可以使用 Boost MPI,它内置了对 Boost Serialization 的支持,而 Boost Serialization 又具有 built-in support for multiprecision.

所以我建议使用它,让自己免于头痛。

简而言之,你已经不在C区了。如果您使用的是 C++,则无法做出 C 程序员倾向于做出的假设。这是最好的,因为您也不需要再编写未定义的行为,或手动完成所有繁琐的工作:)

提升 MPI 演示:

#include <boost/mpi.hpp>
#include <boost/serialization/vector.hpp>
#include <boost/multiprecision/cpp_bin_float.hpp>
#include <boost/multiprecision/gmp.hpp>
#include <iostream>
#include <random>

namespace bmp = boost::multiprecision;
namespace mpi = boost::mpi;

//using Database = std::vector<bmp::mpf_float_50>;
using Database = std::vector<bmp::cpp_bin_float_50>;
static std::mt19937 prng { std::random_device{}() };

int main() {
    mpi::environment env;
    mpi::communicator world;

    if (world.rank() == 0) {
        std::cout << "before for rank=" << world.rank() << std::endl;

        Database db;

        std::generate_n(
                back_inserter(db),
                prng()%15, 
                [i=0]() mutable { return i++*0.1; });

        world.send(1, 1, db);
    } else {
        Database db;
        world.recv(0, 1, db);

        std::cout << "Received " << db.size() << " numbers: ";
        for (auto& number : db) {
            std::cout << number << " ";
        }
        std::cout << std::endl;
    }
}

当运行:

使用不可序列化类型

如果您坚持使用 mpfr 类型,我认为您将不得不手动进行序列化。一种非常天真的方法就是将所有元素转换为字符串:

#include <boost/mpi.hpp>
#include <boost/serialization/vector.hpp>
#include <boost/multiprecision/cpp_bin_float.hpp>
#include <boost/multiprecision/gmp.hpp>
#include <boost/convert/lexical_cast.hpp>
#include <boost/convert.hpp>
#include <boost/range/algorithm.hpp>
#include <boost/range.hpp>
#include <iostream>
#include <random>

namespace bmp = boost::multiprecision;
namespace mpi = boost::mpi;

using Num = bmp::mpf_float_50;
using Database = std::vector<Num>;
using SerializationFormat = std::vector<std::string>;
static auto serialize   = boost::cnv::apply<std::string>(boost::cnv::lexical_cast());
static auto deserialize = boost::cnv::apply<Num>(boost::cnv::lexical_cast());

//using Database = std::vector<bmp::cpp_bin_float_50>;
static std::mt19937 prng { std::random_device{}() };

int main() {
    mpi::environment env;
    mpi::communicator world;

    if (world.rank() == 0) {
        std::cout << "before for rank=" << world.rank() << std::endl;

        Database db;

        std::generate_n(
                back_inserter(db),
                prng()%15, 
                [i=0]() mutable { return i++*0.1; });

        SerializationFormat db_str;
        boost::transform(db, back_inserter(db_str), serialize);
        world.send(1, 1, db_str);
    } else {
        SerializationFormat db_str;
        world.recv(0, 1, db_str);

        Database db;
        boost::transform(db_str, back_inserter(db), deserialize);

        std::cout << "Received " << db.size() << " numbers: ";
        for (auto& number : db) {
            std::cout << number << " ";
        }
        std::cout << std::endl;
    }
}