Question

我正在编写一个用于并行排序的mpich程序。我需要使用mpi_gather接口，但是它不支持传递对象向量。所以我使用 boost_serialization 。

实施

我使用 boost_serialization 序列化载体。

std::string serial_str;
boost::iostreams::back_insert_device<std::string> inserter(serial_str);
boost::iostreams::stream<boost::iostreams::back_insert_device<std::string>> s(inserter);
boost::archive::binary_oarchive send_ar(s);
//samples is the vector<object>
send_ar << samples;
s.flush();
int len = serial_str.size();

然后，我使用 mpi_gather 将所有 serial_str 发送到根进程（ data_recv ）。

char *data_recv = NULL;
if(myid == 0){
data_recv = (char*)malloc(sizeof(char) * (len_all+1));
    data_recv[len_all] = '\0';
}
MPI_Gather((void*)serial_str.data(), len, MPI_BYTE, data_recv, len, MPI_BYTE, 0, MPI_COMM_WORLD);

最后，我反序列化data_recv中的数据。

boost::iostreams::basic_array_source<char> device(data_recv,len_all);
    boost::iostreams::stream<boost::iostreams::basic_array_source<char>> s(device);
    boost::archive::binary_iarchive recv_ar(s);

std::vector<mdata> recv_vec;    
    recv_ar >> recv_vec;

我的实现基于How to send a set object in MPI_Send

问题

我无法正确反序列化data_recv中的数据。我打印了data_recv，然后发现 data_recv 中的数据在mpi_gather之后格式不正确。 第二个存档覆盖了第一个存档。（以粗体显示）

serialization :: archive

XylvXe-M X 000000000000000000000000000020002 DDDDFFFFCCCCBBBBBB111133332222DDDD333388888888FFFF2222 序列化::归档 000000000000000000023D0 EEEE7777EEEE44447777BBBB8888AAAA0000AAAAAAAAFFFF1111

XylvXe-M X 000000000000000000000000000020002 DDDDFFFFCCCCBBBBBB111133332222DDDD333388888888FFFF2222O @ f＆！o，t.b X 000000000000000000000000000023D0 EEEE7777EEEE44447777BBBB8888AAAA0000AAAAAAAAFFFF1111

正确的格式应为：（没有重叠，因此我可以反序列化）

serialization :: archive

XylvXe-M X 000000000000000000000000000020002 DDDDFFFFCCCCBBBBBB111133332222DDDD333388888888FFFF2222O @ f＆！o，t.b X 000000000000000000000000000023D0 EEEE7777EEEE44447777BBBB8888AAAA0000AAAAAAAAFFFF1111

serialization :: archive

XylvXe-M X 000000000000000000000000000020002 DDDDFFFFCCCCBBBBBB111133332222DDDD333388888888FFFF2222O @ f＆！o，t.b X 000000000000000000000000000023D0 EEEE7777EEEE44447777BBBB8888AAAA0000AAAAAAAAFFFF1111

问题

为什么会这样？是否因为mpi_gather与c ++对象不兼容？如果有人可以帮助我，那将解决我的大问题。谢谢！

代码

//processor rank, and total number of processors
int myid, world_size;

//for timing used by root processor
double startwtime = 0.0, endwtime;

//init MPI World
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &world_size);
MPI_Comm_rank(MPI_COMM_WORLD, &myid);
//get the processor name
char processor_name[MPI_MAX_PROCESSOR_NAME];
int name_len;
MPI_Get_processor_name(processor_name,&name_len);

//read local data
std::vector<mdata> mdatas; 
string data_path = "/home/jiang/mpi_data";
readAsciiData(data_path, mdatas);
cout <<"rank: "<<myid <<" mdata_vector.size(): "<<mdatas.size()<<endl;

//local sort according ASCII order 
std::sort(mdatas.begin(), mdatas.end());

//regular sample
std::vector<mdata> samples;
for(int i=0; i<mdatas.size(); i=i+mdatas.size()/world_size){
samples.push_back(mdatas[i]);
}

//gather the regular samples
//passing data in byte stream by using boost serialization

std::string serial_str;
boost::iostreams::back_insert_device<std::string> inserter(serial_str);
boost::iostreams::stream<boost::iostreams::back_insert_device<std::string>> s(inserter);
boost::archive::binary_oarchive send_ar(s);

send_ar << samples;
s.flush();
int len = serial_str.size();
//int len = s.str().size();
int *pivot_lens = NULL;
if(myid == 0){
pivot_lens = (int*)malloc(sizeof(int) * world_size);
}

cout  <<serial_str <<endl;
//first, gathering the lens and calculate the sum
cout << "rank " << myid << " on "<< processor_name << " is sending len: "<< len << endl;
MPI_Gather(&len, 1, MPI_INT, pivot_lens, 1, MPI_INT, 0, MPI_COMM_WORLD);
//calculate the sum of lens
int len_all = 0;
if(myid == 0){
for(int i=0;i<world_size;i++){
       len_all = len_all + pivot_lens[i];
       //cout << pivot_lens[i] << endl;     
}
    cout << "len_all:" << len_all << endl;
    free(pivot_lens);   
}
//then, gathering string of bytes from all the processes
char *data_recv = NULL;
if(myid == 0){
data_recv = (char*)malloc(sizeof(char) * (len_all+1));
    data_recv[len_all] = '\0';

}


MPI_Gather((void*)serial_str.data(), len, MPI_BYTE, data_recv, len, MPI_BYTE, 0, MPI_COMM_WORLD);

// cout << serial_str <<endl;
if(myid == 0){
//deconstructe from byte of string to vector<mdata>
boost::iostreams::basic_array_source<char> device(data_recv,len_all);
    boost::iostreams::stream<boost::iostreams::basic_array_source<char>> s(device);
    boost::archive::binary_iarchive recv_ar(s);

std::vector<mdata> recv_vec;    
    recv_ar >> recv_vec;
int count =0;
for(int i=0;i<len_all;i++){
    cout<<data_recv[i];
    count ++;
}
cout <<endl <<count ;
cout <<endl;
//cout << "rank " << myid << " gets the samples: " << recv_vec.size()<<endl; 
iterateForTest(myid, recv_vec);

free(data_recv);
}

MPI_Finalize();
return 0;

MPI_对象的聚集向量

实施

问题

问题

代码

0 个答案: