MPI并行程序编写与调试（奇偶变换排序）

一、编写MPI并行计算程序

奇偶排序在多线程进行计算代码如下：
关于奇偶排序这里不做过多描述。

/** File:     mpi_odd_even.c* Purpose:  Implement parallel odd-even sort of an array of *           nonegative ints* Input:*    A:     elements of array (optional)* Output:*    A:     elements of A after sorting** Compile:  mpicc -g -Wall -o mpi_odd_even mpi_odd_even.c* Run:*    mpiexec -n <p> mpi_odd_even <g|i> <global_n> *       - p: the number of processes*       - g: generate random, distributed list*       - i: user will input list on process 0*       - global_n: number of elements in global list** Notes:* 1.  global_n must be evenly divisible by p* 2.  Except for debug output, process 0 does all I/O* 3.  Optional -DDEBUG compile flag for verbose output*/
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <mpi.h>const int RMAX = 100;/* Local functions */
void Usage(char* program);
void Print_list(int local_A[], int local_n, int rank);
void Merge_low(int local_A[], int temp_B[], int temp_C[], int local_n);
void Merge_high(int local_A[], int temp_B[], int temp_C[], int local_n);
void Generate_list(int local_A[], int local_n, int my_rank);
int  Compare(const void* a_p, const void* b_p);/* Functions involving communication */
void Get_args(int argc, char* argv[], int* global_n_p, int* local_n_p, char* gi_p, int my_rank, int p, MPI_Comm comm);
void Sort(int local_A[], int local_n, int my_rank, int p, MPI_Comm comm);
void Odd_even_iter(int local_A[], int temp_B[], int temp_C[],int local_n, int phase, int even_partner, int odd_partner,int my_rank, int p, MPI_Comm comm);
void Print_local_lists(int local_A[], int local_n, int my_rank, int p, MPI_Comm comm);
void Print_global_list(int local_A[], int local_n, int my_rank,int p, MPI_Comm comm);
void Read_list(int local_A[], int local_n, int my_rank, int p,MPI_Comm comm);/*-------------------------------------------------------------------*/
int main(int argc, char* argv[]) {int my_rank, p;char g_i;int *local_A;int global_n;int local_n;MPI_Comm comm;MPI_Init(&argc, &argv);comm = MPI_COMM_WORLD;MPI_Comm_size(comm, &p);MPI_Comm_rank(comm, &my_rank);Get_args(argc, argv, &global_n, &local_n, &g_i, my_rank, p, comm);local_A = (int*) malloc(local_n*sizeof(int));if (g_i == 'g') {Generate_list(local_A, local_n, my_rank);Print_local_lists(local_A, local_n, my_rank, p, comm);} else {Read_list(local_A, local_n, my_rank, p, comm);
#     ifdef DEBUGPrint_local_lists(local_A, local_n, my_rank, p, comm);
#     endif}#  ifdef DEBUGprintf("Proc %d > Before Sort\n", my_rank);fflush(stdout);
#  endifSort(local_A, local_n, my_rank, p, comm);#  ifdef DEBUGPrint_local_lists(local_A, local_n, my_rank, p, comm);fflush(stdout);
#  endifPrint_global_list(local_A, local_n, my_rank, p, comm);free(local_A);MPI_Finalize();return 0;
}  /* main *//*-------------------------------------------------------------------* Function:   Generate_list* Purpose:    Fill list with random ints* Input Args: local_n, my_rank* Output Arg: local_A*/
void Generate_list(int local_A[], int local_n, int my_rank) {int i;srandom(my_rank+1);for (i = 0; i < local_n; i++)local_A[i] = random() % RMAX;}  /* Generate_list *//*-------------------------------------------------------------------* Function:  Usage* Purpose:   Print command line to start program* In arg:    program:  name of executable* Note:      Purely local, run only by process 0;*/
void Usage(char* program) {fprintf(stderr, "usage:  mpirun -np <p> %s <g|i> <global_n>\n",program);fprintf(stderr, "   - p: the number of processes \n");fprintf(stderr, "   - g: generate random, distributed list\n");fprintf(stderr, "   - i: user will input list on process 0\n");fprintf(stderr, "   - global_n: number of elements in global list");fprintf(stderr, " (must be evenly divisible by p)\n");fflush(stderr);
}  /* Usage *//*-------------------------------------------------------------------* Function:    Get_args* Purpose:     Get and check command line arguments* Input args:  argc, argv, my_rank, p, comm* Output args: global_n_p, local_n_p, gi_p*/
void Get_args(int argc, char* argv[], int* global_n_p, int* local_n_p, char* gi_p, int my_rank, int p, MPI_Comm comm) {if (my_rank == 0) {if (argc != 3) {Usage(argv[0]);*global_n_p = -1;  /* Bad args, quit */} else {*gi_p = argv[1][0];if (*gi_p != 'g' && *gi_p != 'i') {Usage(argv[0]);*global_n_p = -1;  /* Bad args, quit */} else {*global_n_p = atoi(argv[2]);if (*global_n_p % p != 0) {Usage(argv[0]);*global_n_p = -1;}}}}  /* my_rank == 0 */MPI_Bcast(gi_p, 1, MPI_CHAR, 0, comm);MPI_Bcast(global_n_p, 1, MPI_INT, 0, comm);if (*global_n_p <= 0) {MPI_Finalize();exit(-1);}*local_n_p = *global_n_p/p;
#  ifdef DEBUGprintf("Proc %d > gi = %c, global_n = %d, local_n = %d\n",my_rank, *gi_p, *global_n_p, *local_n_p);fflush(stdout);
#  endif}  /* Get_args *//*-------------------------------------------------------------------* Function:   Read_list* Purpose:    process 0 reads the list from stdin and scatters it*             to the other processes.* In args:    local_n, my_rank, p, comm* Out arg:    local_A*/
void Read_list(int local_A[], int local_n, int my_rank, int p,MPI_Comm comm) {int i;int *temp;if (my_rank == 0) {temp = (int*) malloc(p*local_n*sizeof(int));printf("Enter the elements of the list\n");for (i = 0; i < p*local_n; i++)scanf("%d", &temp[i]);} MPI_Scatter(temp, local_n, MPI_INT, local_A, local_n, MPI_INT,0, comm);if (my_rank == 0)free(temp);
}  /* Read_list *//*-------------------------------------------------------------------* Function:   Print_global_list* Purpose:    Print the contents of the global list A* Input args:  *    n, the number of elements *    A, the list* Note:       Purely local, called only by process 0*/
void Print_global_list(int local_A[], int local_n, int my_rank, int p, MPI_Comm comm) {int* A;int i, n;if (my_rank == 0) {n = p*local_n;A = (int*) malloc(n*sizeof(int));MPI_Gather(local_A, local_n, MPI_INT, A, local_n, MPI_INT, 0,comm);printf("Global list:\n");for (i = 0; i < n; i++)printf("%d ", A[i]);printf("\n\n");free(A);} else {MPI_Gather(local_A, local_n, MPI_INT, A, local_n, MPI_INT, 0,comm);}}  /* Print_global_list *//*-------------------------------------------------------------------* Function:    Compare* Purpose:     Compare 2 ints, return -1, 0, or 1, respectively, when*              the first int is less than, equal, or greater than*              the second.  Used by qsort.*/
int Compare(const void* a_p, const void* b_p) {int a = *((int*)a_p);int b = *((int*)b_p);if (a < b)return -1;else if (a == b)return 0;else /* a > b */return 1;
}  /* Compare *//*-------------------------------------------------------------------* Function:    Sort* Purpose:     Sort local list, use odd-even sort to sort*              global list.* Input args:  local_n, my_rank, p, comm* In/out args: local_A */
void Sort(int local_A[], int local_n, int my_rank, int p, MPI_Comm comm) {int phase;int *temp_B, *temp_C;int even_partner;  /* phase is even or left-looking */int odd_partner;   /* phase is odd or right-looking *//* Temporary storage used in merge-split */temp_B = (int*) malloc(local_n*sizeof(int));temp_C = (int*) malloc(local_n*sizeof(int));/* Find partners:  negative rank => do nothing during phase */if (my_rank % 2 != 0) {even_partner = my_rank - 1;odd_partner = my_rank + 1;if (odd_partner == p) odd_partner = MPI_PROC_NULL;  // Idle during odd phase} else {even_partner = my_rank + 1;if (even_partner == p) even_partner = MPI_PROC_NULL;  // Idle during even phaseodd_partner = my_rank-1;  }/* Sort local list using built-in quick sort */qsort(local_A, local_n, sizeof(int), Compare);#  ifdef DEBUGprintf("Proc %d > before loop in sort\n", my_rank);fflush(stdout);
#  endiffor (phase = 0; phase < p; phase++)Odd_even_iter(local_A, temp_B, temp_C, local_n, phase, even_partner, odd_partner, my_rank, p, comm);free(temp_B);free(temp_C);
}  /* Sort *//*-------------------------------------------------------------------* Function:    Odd_even_iter* Purpose:     One iteration of Odd-even transposition sort* In args:     local_n, phase, my_rank, p, comm* In/out args: local_A* Scratch:     temp_B, temp_C*/
void Odd_even_iter(int local_A[], int temp_B[], int temp_C[],int local_n, int phase, int even_partner, int odd_partner,int my_rank, int p, MPI_Comm comm) {MPI_Status status;if (phase % 2 == 0) {if (even_partner >= 0) {MPI_Sendrecv(local_A, local_n, MPI_INT, even_partner, 0, temp_B, local_n, MPI_INT, even_partner, 0, comm,&status);if (my_rank % 2 != 0)Merge_high(local_A, temp_B, temp_C, local_n);elseMerge_low(local_A, temp_B, temp_C, local_n);}} else { /* odd phase */if (odd_partner >= 0) {MPI_Sendrecv(local_A, local_n, MPI_INT, odd_partner, 0, temp_B, local_n, MPI_INT, odd_partner, 0, comm,&status);if (my_rank % 2 != 0)Merge_low(local_A, temp_B, temp_C, local_n);elseMerge_high(local_A, temp_B, temp_C, local_n);}}
}  /* Odd_even_iter *//*-------------------------------------------------------------------* Function:    Merge_low* Purpose:     Merge the smallest local_n elements in my_keys*              and recv_keys into temp_keys.  Then copy temp_keys*              back into my_keys.* In args:     local_n, recv_keys* In/out args: my_keys* Scratch:     temp_keys*/
void Merge_low(int  my_keys[],     /* in/out    */int  recv_keys[],   /* in        */int  temp_keys[],   /* scratch   */int  local_n        /* = n/p, in */) {int m_i, r_i, t_i;m_i = r_i = t_i = 0;while (t_i < local_n) {if (my_keys[m_i] <= recv_keys[r_i]) {temp_keys[t_i] = my_keys[m_i];t_i++; m_i++;} else {temp_keys[t_i] = recv_keys[r_i];t_i++; r_i++;}}memcpy(my_keys, temp_keys, local_n*sizeof(int));
}  /* Merge_low *//*-------------------------------------------------------------------* Function:    Merge_high* Purpose:     Merge the largest local_n elements in local_A *              and temp_B into temp_C.  Then copy temp_C*              back into local_A.* In args:     local_n, temp_B* In/out args: local_A* Scratch:     temp_C*/
void Merge_high(int local_A[], int temp_B[], int temp_C[], int local_n) {int ai, bi, ci;ai = local_n-1;bi = local_n-1;ci = local_n-1;while (ci >= 0) {if (local_A[ai] >= temp_B[bi]) {temp_C[ci] = local_A[ai];ci--; ai--;} else {temp_C[ci] = temp_B[bi];ci--; bi--;}}memcpy(local_A, temp_C, local_n*sizeof(int));
}  /* Merge_high *//*-------------------------------------------------------------------* Only called by process 0*/
void Print_list(int local_A[], int local_n, int rank) {int i;printf("%d: ", rank);for (i = 0; i < local_n; i++)printf("%d ", local_A[i]);printf("\n");
}  /* Print_list *//*-------------------------------------------------------------------* Function:   Print_local_lists* Purpose:    Print each process' current list contents* Input args: all* Notes:* 1.  Assumes all participating processes are contributing local_n *     elements*/
void Print_local_lists(int local_A[], int local_n, int my_rank, int p, MPI_Comm comm) {int*       A;int        q;MPI_Status status;if (my_rank == 0) {A = (int*) malloc(local_n*sizeof(int));Print_list(local_A, local_n, my_rank);for (q = 1; q < p; q++) {MPI_Recv(A, local_n, MPI_INT, q, 0, comm, &status);Print_list(A, local_n, q);}free(A);} else {MPI_Send(local_A, local_n, MPI_INT, 0, 0, comm);}
}  /* Print_local_lists */

二、编译

三、多次运行结果

四、结论

一是在写代码调用多个核心多线程运算时，声明变量需编写在前面，否则会造成编译失败；二是按照实验原理来说，多核运算的运算时间肯定是要小于单核运算的，但是在很少的情况下，多核运算的耗时可能反而高于单核运算（如上图），这是因为如上图的情况下，比如生成100个随机数，不是相同的100个随机数，很可能多核运算需要排序的100个数字本身就需要更多的判断排序，所以可能会造成多核运算的耗时高，但是随着排序数量的增加，其多核运算的耗时明显会和单核拉开距离，所以如果数据量本身就少，可能会有个例，但是随着数据量的增加，多核运算的耗时明显会低于单核运算，通过对比图中排序数量高于200的耗时对比，说明理论是正确的，同时也要注意核心的分配，不要分配比硬件本身更高的核心，这样反而会降低速度。

MPI并行程序编写与调试（奇偶变换排序）相关推荐

c语言mpi并行程序,高性能计算之并行编程技术MPI并行程序设计(完整版).pdf
高性能计算之并行编程技术MPI并行程序设计(完整版) 高性能计算之并行编程技术 -- MPI并行程序设计都志辉编著李三立审阅陈渝刘鹏校对 I 内容提要本书介绍目前最常见的并行程序- M ...
简单stm32程序编写以及调试
简单stm32程序编写以及调试一.环境配置 1.MDK的安装 MDK(Microcontroller Development Kit)是针对ARM处理器,特别是Cortex-M内核处理器的最佳开发工 ...
MPI并行程序的调试技巧
原文地址:http://galoisplusplus.gitcafe.com/blog/2013/06/08/mpi-debug-tips/ debug一个并行程序(parallel program) ...
c语言mpi并行程序,[转]MPI并行程序的调试技巧
debug一个并行程序(parallel program)向来是件很麻烦的事情(Erlang等functional programming language另当别论),对于像MPI这种非shared ...
MPI并行程序开发设计----------------------------------并行编程模型和算法等介绍
---------------------------------------------------------------------------------------------------- ...
Linux系统上C语言程序编写与调试
目录 1.可执行文件(程序) 2.程序的编译链接过程 2.1.预处理阶段 2.2.编译阶段 2.3.汇编阶段 2.4.链接阶段 3.在gcc编译器上进行程序编译运行实操 3.1.gcc安装 3.2.预 ...
【MPI学习3】MPI并行程序设计模式：不同通信模式MPI并行程序的设计
学习了MPI四种通信模式及其函数用法: (1)标准通信模式:MPI_SEND (2)缓存通信模式:MPI_BSEND (3)同步通信模式:MPI_SSEND (4)就绪通信模式:MPI_RSEND ...
MPI并行程序-矩阵分块乘法( Matrix multiplication : Two-Demension Method)
问题引入: 主体思路:(为了尽可能使计算简便,所有的实验矩阵均是方阵) 1. 对于A,B矩阵,首先根据子进程数p将其划分成p块形式相同的小矩阵块,其中每个矩阵块只要求是方阵即可,具体可以是1阶的 ...
c语言mpi矩阵乘法,【MPI并行程序】矩阵乘法
具体思路为创建矩阵,给每个进程分配矩阵的不同部分,每个进程根据进程号的不同处理矩阵的不同部分,计算完成后将每个进程的数组发送给一个特定的进程,由该进程合成一个完整的矩阵输出出来.问题是不一定每个进程都 ...

MPI并行程序编写与调试（奇偶变换排序）

MPI并行程序编写与调试（奇偶变换排序）

一、编写MPI并行计算程序

二、编译

三、多次运行结果

四、结论

MPI并行程序编写与调试（奇偶变换排序）相关推荐

最新文章

热门文章