创建动态大小的MPI文件视图 [英] Creating dynamically sized MPI file views
问题描述
我想使用集体MPI I/O写出一个二进制文件.我的计划是创建一个类似于
I would like to write out a binary file using collective MPI I/O. My plan is to create an MPI derived type analogous to
struct soln_dynamic_t
{
int int_data[2];
double *u; /* Length constant for all instances of this struct */
};
然后，每个处理器都基于派生的类型创建一个视图，并将其写入该视图.
Each processor then creates a view based on the derived type, and writes into the view.
对于用u[10]替换*u的情况，我都可以使用(请参见下面的完整代码)，但是最终，我想为u提供一个动态长度数组. (如果重要的话，soln_dynamic_t的所有实例的长度将是固定的，但是在编译时未知.)
I have this all working for the case in which *u is replaced with u[10] (see complete code below), but ultimately, I'd like to have a dynamic length array for u. (In case it matters, the length will be fixed for all instances of soln_dynamic_t for any run, but not known at compile time.)
处理此问题的最佳方法是什么?
What is the best way to handle this?
我已经阅读了几篇关于为什么不能使用soln_dynamic_t的文章
直接作为MPI结构.问题是不能保证处理器在u[0]和int_data[0]之间具有相同的偏移量. (是吗?)
I have read several posts on why I can't use soln_dynamic_t
directly as an MPI structure. The problem is that processors are not guaranteed to have the same offset between u[0] and int_data[0]. (Is that right?)
另一方面，结构
struct soln_static_t
{
int int_data[2];
double u[10]; /* fixed at compile time */
};
之所以起作用，是因为偏移量在各个处理器之间保证是相同的.
works because the offsets are guaranteed to be the same across processors.
我考虑了几种方法:
-
基于手动定义的偏移量等来创建视图，而不是使用派生类型.
Create the view based on manually defined offsets, etc, rather than using a derived type.
基于另一个MPI类型的MPI结构，即``* u`(允许吗?)的连续类型
Base the MPI structure on another MPI type, i.e. an contiguous type for ``*u` (is that allowed?)
我猜想必须有一种标准的方法来做到这一点.任何建议都将非常有帮助.
I am guessing there must be a standard way to do this. Any suggestions would be very helpful.
关于此问题的其他几篇文章也很有帮助，尽管它们主要处理通讯问题，而不是文件I/O.
Several other posts on this issue have been helpful, although they mostly deal with communication and not file I/O.
这是完整的代码::
#include <mpi.h>
typedef struct
{
int int_data[2];
double u[10]; /* Make this a dynamic length (but fixed) */
} soln_static_t;
void build_soln_type(int n, int* int_data, double *u, MPI_Datatype *soln_t)
{
int block_lengths[2] = {2,n};
MPI_Datatype typelist[2] = {MPI_INT, MPI_DOUBLE};
MPI_Aint disp[2], start_address, address;
MPI_Address(int_data,&start_address);
MPI_Address(u,&address);
disp[0] = 0;
disp[1] = address-start_address;
MPI_Datatype tmp_type;
MPI_Type_create_struct(2,block_lengths,disp,typelist,&tmp_type);
MPI_Aint extent;
extent = block_lengths[0]*sizeof(int) + block_lengths[1]*sizeof(double);
MPI_Type_create_resized(tmp_type, 0, extent, soln_t);
MPI_Type_commit(soln_t);
}
void main(int argc, char** argv)
{
MPI_File file;
int globalsize, localsize, starts, order;
MPI_Datatype localarray, soln_t;
int rank, nprocs, nsize = 10; /* must match size in struct above */
/* --- Initialize MPI */
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
/* --- Set up data to write out */
soln_static_t data;
data.int_data[0] = nsize;
data.int_data[1] = rank;
data.u[0] = 3.14159; /* To check that data is written as expected */
build_soln_type(nsize, data.int_data, data.u, &soln_t);
MPI_File_open(MPI_COMM_WORLD, "bin.out",
MPI_MODE_CREATE|MPI_MODE_WRONLY,
MPI_INFO_NULL, &file);
/* --- Create file view for this processor */
globalsize = nprocs;
localsize = 1;
starts = rank;
order = MPI_ORDER_C;
MPI_Type_create_subarray(1, &globalsize, &localsize, &starts, order,
soln_t, &localarray);
MPI_Type_commit(&localarray);
MPI_File_set_view(file, 0, soln_t, localarray,
"native", MPI_INFO_NULL);
/* --- Write data into view */
MPI_File_write_all(file, data.int_data, 1, soln_t, MPI_STATUS_IGNORE);
/* --- Clean up */
MPI_File_close(&file);
MPI_Type_free(&localarray);
MPI_Type_free(&soln_t);
MPI_Finalize();
}
推荐答案
由于soln_dynamic_t类型的u数组的大小在运行时是已知的，此后将不会更改，因此我建议使用其他方法.方法.
Since the size of the u array of the soln_dynamic_t type is known at runtime and will not change after that, I'd rather suggest an other approach.
基本上，您将所有连续的数据存储在内存中:
Basically, you store all the data contiguous in memory :
typedef struct
{
int int_data[2];
double u[]; /* Make this a dynamic length (but fixed) */
} soln_dynamic_t;
然后您必须手动分配此结构
Then you have to manually allocate this struct
soln_dynamic_t * alloc_soln(int nsize, int count) {
return (soln_dynamic_t *)calloc(offsetof(soln_dynamic_t, u)+nsize*sizeof(double), count);
}
请注意，由于在编译时大小未知，因此无法直接访问soln_dynamic_t数组.相反，您必须手动计算指针.
Note you cannot directly access an array of soln_dynamic_t because the size is unknown at compile time. Instead, you have to manually calculate the pointers.
soln_dynamic_t *p = alloc_soln(10, 2);
p[0].int_data[0] = 1; // OK
p[0].u[0] = 2; // OK
p[1].int_data[0] = 3; // KO ! since sizeof(soln_dynamic_t) is unknown at compile time.
这是程序的完整重写版本
Here is the full rewritten version of your program
#include <mpi.h>
#include <malloc.h>
typedef struct
{
int int_data[2];
double u[]; /* Make this a dynamic length (but fixed) */
} soln_dynamic_t;
void build_soln_type(int n, MPI_Datatype *soln_t)
{
int block_lengths[2] = {2,n};
MPI_Datatype typelist[2] = {MPI_INT, MPI_DOUBLE};
MPI_Aint disp[2];
disp[0] = offsetof(soln_dynamic_t, int_data);
disp[1] = offsetof(soln_dynamic_t, u);
MPI_Datatype tmp_type;
MPI_Type_create_struct(2,block_lengths,disp,typelist,&tmp_type);
MPI_Aint extent;
extent = offsetof(soln_dynamic_t, u) + block_lengths[1]*sizeof(double);
MPI_Type_create_resized(tmp_type, 0, extent, soln_t);
MPI_Type_free(&tmp_type);
MPI_Type_commit(soln_t);
}
soln_dynamic_t * alloc_soln(int nsize, int count) {
return (soln_dynamic_t *)calloc(offsetof(soln_dynamic_t, u) + nsize*sizeof(double), count);
}
int main(int argc, char** argv)
{
MPI_File file;
int globalsize, localsize, starts, order;
MPI_Datatype localarray, soln_t;
int rank, nprocs, nsize = 10; /* must match size in struct above */
/* --- Initialize MPI */
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
/* --- Set up data to write out */
soln_dynamic_t *data = alloc_soln(nsize,1);
data->int_data[0] = nsize;
data->int_data[1] = rank;
data->u[0] = 3.14159; /* To check that data is written as expected */
build_soln_type(nsize, &soln_t);
MPI_File_open(MPI_COMM_WORLD, "bin2.out",
MPI_MODE_CREATE|MPI_MODE_WRONLY,
MPI_INFO_NULL, &file);
/* --- Create file view for this processor */
globalsize = nprocs;
localsize = 1;
starts = rank;
order = MPI_ORDER_C;
MPI_Type_create_subarray(1, &globalsize, &localsize, &starts, order,
soln_t, &localarray);
MPI_Type_commit(&localarray);
MPI_File_set_view(file, 0, soln_t, localarray,
"native", MPI_INFO_NULL);
/* --- Write data into view */
MPI_File_write_all(file, data, 1, soln_t, MPI_STATUS_IGNORE);
/* --- Clean up */
MPI_File_close(&file);
MPI_Type_free(&localarray);
MPI_Type_free(&soln_t);
MPI_Finalize();
return 0;
}
这篇关于创建动态大小的MPI文件视图的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！
