未知子 class 与 boost::mpi::packed_oarchive 和 packed_iarchive 的 MPI 传输

MPI transmission of unknown sub-class with boost::mpi::packed_oarchive and packed_iarchive

我正在尝试传输一个 class 的未知子 class,但已知基 class。

我相信这应该可以使用 boost::serializationBOOST_CLASS_EXPORT_GUIDboost::mpi,但总的来说我对 C++ 还很陌生

这是我的代码:

#include <boost/mpi.hpp>
#include <boost/mpi/environment.hpp>
#include <boost/mpi/communicator.hpp>
#include <boost/serialization/string.hpp>
#include <boost/serialization/serialization.hpp>
#include <boost/serialization/export.hpp>
#include <boost/serialization/vector.hpp>
#include <iostream>

namespace mpi = boost::mpi;


class Action {
protected:
    int start_rank;
    std::string greeting;

    Action(std::string greeting) {
        mpi::communicator world;
        this->start_rank = world.rank();
        this->greeting = greeting;
    };

private:

    friend class boost::serialization::access;
    template<class Archive> void serialize(Archive &ar, const unsigned int version) {
        ar & this->start_rank;
        ar & this->greeting;
    };

public:

    Action() = default;

    void invoke() {
        mpi::communicator world;
        std::cout << this->greeting << "! I am process " << world.rank() << " of " << world.size()
            << ". I was created on " << this->start_rank << "." << std::endl;
    };
};


class HelloAction : public Action {

public:
    HelloAction() : Action("Hello") {};

};

class GoodByeAction : public Action {

public:
    GoodByeAction() : Action("Good bye") {};

};

BOOST_CLASS_EXPORT_GUID(Action, "Action");
BOOST_CLASS_EXPORT_GUID(HelloAction, "HelloAction");
BOOST_CLASS_EXPORT_GUID(GoodByeAction, "GoodByeAction");

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

    HelloAction *hello = new HelloAction();
    mpi::broadcast(world, hello, 0);
    hello->invoke();

    GoodByeAction *bye = new GoodByeAction();
    mpi::broadcast(world, bye, 1);
    bye->invoke();

    world.barrier();

    if (world.rank() == 0) {
        std::cout << "sending unknown action classes!" << std::endl;
        HelloAction *yup = new HelloAction();
        boost::mpi::packed_oarchive oar(world);
        oar << yup;
    }
    else {
        std::cout << "receiving unknown action classes!" << std::endl;
        Action *action = NULL;
        boost::mpi::packed_iarchive iar(world);
        iar >> action;
        action->invoke();
    }

    return 0;
}

compiling/running 与:

mpic++ -g -std=c++1y hello.cpp -lboost_serialization -lmpi -lboost_mpi
mpiexec -np 2 ./a.out

这个好像运行就好了:

Hello! I am process 0 of 2. I was created on 0.
Hello! I am process 1 of 2. I was created on 0.
Good bye! I am process 1 of 2. I was created on 1.
Good bye! I am process 0 of 2. I was created on 1.

... 直到到达 "sending/receiving of unknown action classes",在那里我得到 运行 时间错误:

receiving unknown action classes!
sending unknown action classes!
terminate called after throwing an instance of 'boost::exception_detail::clone_impl<boost::exception_detail::error_info_injector<boost::mpi::exception> >'
  what():  MPI_Unpack: MPI_ERR_ARG: invalid argument of some other kind
[machine-name:20194] *** Process received signal ***
[machine-name:20194] Signal: Aborted (6)
[machine-name:20194] Signal code:  (-6)
[machine-name:20194] [ 0] /lib/x86_64-linux-gnu/libc.so.6(+0x352f0) [0x7fa685cc22f0]
[machine-name:20194] [ 1] /lib/x86_64-linux-gnu/libc.so.6(gsignal+0x37) [0x7fa685cc2267]
[machine-name:20194] [ 2] /lib/x86_64-linux-gnu/libc.so.6(abort+0x16a) [0x7fa685cc3eca]
[machine-name:20194] [ 3] /usr/lib/x86_64-linux-gnu/libstdc++.so.6(_ZN9__gnu_cxx27__verbose_terminate_handlerEv+0x16d) [0x7fa6862fdb7d]
[machine-name:20194] [ 4] /usr/lib/x86_64-linux-gnu/libstdc++.so.6(+0x8d9c6) [0x7fa6862fb9c6]
[machine-name:20194] [ 5] /usr/lib/x86_64-linux-gnu/libstdc++.so.6(+0x8da11) [0x7fa6862fba11]
[machine-name:20194] [ 6] /usr/lib/x86_64-linux-gnu/libstdc++.so.6(+0x8dc29) [0x7fa6862fbc29]
[machine-name:20194] [ 7] ./a.out(_ZN5boost15throw_exceptionINS_3mpi9exceptionEEEvRKT_+0x80) [0x41426d]
[machine-name:20194] [ 8] ./a.out() [0x413802]
[machine-name:20194] [15] ./a.out() [0x4306e6]
[machine-name:20194] [16] /usr/lib/x86_64-linux-gnu/libboost_serialization.so.1.58.0(_ZN5boost7archive6detail19basic_iarchive_impl12load_pointerERNS1_14basic_iarchiveERPvPKNS1_25basic_pointer_iserializerEPFS9_RKNS_13serialization18extended_type_infoEE+0x4d) [0x7fa686df5b8d]
[machine-name:20194] [17] ./a.out() [0x41d08d]
[machine-name:20194] [23] ./a.out() [0x40d293]
[machine-name:20194] [24] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf0) [0x7fa685cada40]
[machine-name:20194] [25] ./a.out() [0x40cfc9]
[machine-name:20194] *** End of error message ***
--------------------------------------------------------------------------
mpiexec noticed that process rank 1 with PID 20194 on node machine-name exited on signal 6 (Aborted).
--------------------------------------------------------------------------

问题:

  1. 尽管这个问题被标记为 boost,是否有可能在有或没有增强的情况下传输一个未知的 subclass?我知道我可以按字符串名称进行各种注册,以及一个控制反转的人来处理创建,但我认为这是 BOOST_CLASS_EXPORT_GUID.
  2. 的意图
  3. 有什么方法可以让我用我现有的东西工作吗?
  4. 是否有一个相当简单的替代方案(不使用 boost)使它与常规的 ol' MPI 一起工作?

您非常接近:首先,您缺少压缩档案的实际通信功能。您正在尝试从 packed_iarchive 中提取内容,但未将任何内容接收到该存档中。例如,您可以将 broadcast 用作 described:

with packed archives, the root sends a packed_oarchive [...] whereas the other processes receive a acked_iarchive [...].

但是,只需将调用添加到您的示例中,就会实例化一个 Action 而不是 HelloAction。需要更多的努力:

  • 使用非虚拟 classes 的基 class 指针不是一个好主意。它可能适用于您的情况,因为 subclasses 只不过是不同的构造函数,但对于稍微复杂的示例来说它会成为问题。
  • Boost 期望(反)序列化调用使用对称类型。您不能序列化 HelloAction* 然后反序列化 Action*,无论它是否是虚拟 class1。如果你在这两种情况下都必须使用Action*,你当然也必须使用virtual classes,这样boost实际上知道运行时类型。从技术上讲,您可以将一个 Action* 序列化为一个 HelloAction 对象,然后反序列化一个 Action* 以生成原始对象的 sliced Action 部分。同样,这适用于您的示例,但它不适用于更复杂的示例,因此不是一个好主意。
  • 您还必须为子 class 专精 serialize1.

将所有这些放在一起,一个工作示例如下所示: 注意:我添加了更多的输出来验证是否使用正确的类型创建了对象。

#include <boost/mpi.hpp>
#include <boost/mpi/environment.hpp>
#include <boost/mpi/communicator.hpp>
#include <boost/serialization/string.hpp>
#include <boost/serialization/serialization.hpp>
#include <boost/serialization/export.hpp>
#include <boost/serialization/vector.hpp>
#include <iostream>

namespace mpi = boost::mpi;


class Action {
protected:
    int start_rank;
    std::string greeting;

    Action(std::string greeting) {
        mpi::communicator world;
        std::cout << world.rank() << ": Creating a Action: " << greeting << std::endl;
        this->start_rank = world.rank();
        this->greeting = greeting;
    };
private:

    friend class boost::serialization::access;
    template<class Archive> void serialize(Archive &ar, const unsigned int version) {
        ar & this->start_rank;
        ar & this->greeting;
    };

public:

    virtual ~Action() = default;

    Action() { 
        mpi::communicator world;
        std::cout << world.rank() << ": Creating a naked Action" << std::endl;
    }

    void invoke() {
        mpi::communicator world;
        std::cout << this->greeting << "! I am process " << world.rank() << " of " << world.size()
            << ". I was created on " << this->start_rank << "." << std::endl;
    };
};


class HelloAction : public Action {

public:
    HelloAction() : Action("Hello") {
        mpi::communicator world;
        std::cout << world.rank() << ": Creating an HelloAction" << std::endl;
    }

    template<typename Archive>  
    void serialize(Archive & ar, const unsigned int file_version) {  
        ar & boost::serialization::base_object<Action>(*this);
    }
};

class GoodByeAction : public Action {

public:
    GoodByeAction() : Action("Good bye") {
        mpi::communicator world;
        std::cout << world.rank() << ": Creating an GoodByeAction" << std::endl;
    }

    template<typename Archive>  
    void serialize(Archive & ar, const unsigned int file_version) {  
        ar & boost::serialization::base_object<Action>(*this);
    }
};

// It is totally fine to use the EXPORT (without GUID) for MPI
BOOST_CLASS_EXPORT(GoodByeAction)
BOOST_CLASS_EXPORT(HelloAction)

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

    HelloAction *hello = new HelloAction();
    mpi::broadcast(world, hello, 0);
    hello->invoke();

    GoodByeAction *bye = new GoodByeAction();
    mpi::broadcast(world, bye, 1);
    bye->invoke();

    world.barrier();

    if (world.rank() == 0) {
        std::cout << "sending unknown action classes!" << std::endl;
        Action *yup = new HelloAction();
        boost::mpi::packed_oarchive oar(world);
        oar << yup;
        mpi::broadcast(world, oar, 0);
    }
    else {
        std::cout << "receiving unknown action classes!" << std::endl;
        boost::mpi::packed_iarchive iar(world);
        mpi::broadcast(world, iar, 0);
        Action *action = nullptr;
        iar >> action;
        std::cout << "RTTI: " << typeid(*action).name() << std::endl;
        action->invoke();
    }

    return 0;
}

关于普通的旧 MPI:对我来说听起来像是重新发明轮子 - 当然可能,但不一定可取。

1 我在文档中找不到关于此的明确声明。这是基于如果你不这样尝试,事情就会变得非常糟糕。我将不胜感激对此发表评论/更好的回答。