在MPI中如何广播C++向量？

Question

我希望使用MPI广播C++向量。

我不可以使用boost.mpi库。

目前，我正在使用stackoverflow中得票最高的答案 Vector Usage in MPI(C++), 但它不能正常工作。

好的，这是代码：

// declaration of variables (ParsedData object will contain these variables)
int generators_count, intervals_count;
std::vector<float> mean_arr, variance_arr, interval_begins_arr, interval_ends_arr;
std::vector<int> amount_of_numbers_to_generate_arr;

// load data into parsed_data object
if (process_id == 0) {
  ParsedData parsed_data = load_input_data(filename);
  intervals_count = parsed_data.intervals_count;
  generators_count = parsed_data.generators_count;
  mean_arr = parsed_data.mean_arr;
  variance_arr = parsed_data.variance_arr;
  interval_begins_arr = parsed_data.interval_begins_arr;
  interval_ends_arr = parsed_data.interval_ends_arr;
  amount_of_numbers_to_generate_arr = parsed_data.amount_of_numbers_to_generate_arr;
}

// send size of vectors to all processes - that MPI code works
MPI_Bcast(&intervals_count, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(&generators_count, 1, MPI_INT, 0, MPI_COMM_WORLD);

// reserve memory for vectors
if (process_id != 0) {
  mean_arr.reserve(generators_count);
  variance_arr.reserve(generators_count);
  amount_of_numbers_to_generate_arr.reserve(generators_count);

  interval_begins_arr.reserve(intervals_count);
  interval_ends_arr.reserve(intervals_count);
}

// stolen from https://dev59.com/PkzSa4cB1Zd3GeqPqeNG
// broadcast each vector, following code compiles but not works properly
MPI_Bcast(&mean_arr[0], mean_arr.size(), MPI_FLOAT, 0, MPI_COMM_WORLD);
MPI_Bcast(&variance_arr[0], variance_arr.size(), MPI_FLOAT, 0, MPI_COMM_WORLD);
MPI_Bcast(&interval_begins_arr[0], interval_begins_arr.size(), MPI_FLOAT, 0, MPI_COMM_WORLD);
MPI_Bcast(&interval_ends_arr[0], interval_ends_arr.size(), MPI_FLOAT, 0, MPI_COMM_WORLD);
MPI_Bcast(&amount_of_numbers_to_generate_arr[0], amount_of_numbers_to_generate_arr.size(), MPI_FLOAT, 0, MPI_COMM_WORLD);

当我简单地将结果放在不同的进程中时，我得到了错误的答案:

  cout << "process ID " << process_id << endl << "amount_of_numbers_to_generate_arr[0] " << amount_of_numbers_to_generate_arr[0] << endl;

结果：

using process 0
process ID 0
amount_of_numbers_to_generate_arr[0] 100000000
using process 1
process ID 1
amount_of_numbers_to_generate_arr[0] -423989576
using process 2
process ID 2
amount_of_numbers_to_generate_arr[0] 1741864632

并且应该是：

using process 0
process ID 0
amount_of_numbers_to_generate_arr[0] 100000000
using process 1
process ID 1
amount_of_numbers_to_generate_arr[0] 100000000
using process 2
process ID 2
amount_of_numbers_to_generate_arr[0] 100000000

无论如何，如何解决这个问题？或者也许有另一种方法来解决它？

- nothing-special-here

---

这里有一个类似的问题，专门涉及传递结构体向量。链接 - Madeleine P. Vincent

Answer 1

当您“为向量保留内存”时，实际上并没有调整向量的大小。在接下来的MPI_Bcast调用中，向量的大小仍然为0，因此不会接收任何值。

要实际调整向量的大小，请调用resize而不是reserve。

- Mankarse