原本以爲，MPI天生只能在Linux上運行。但這次卻發現了Intel MPI Library 這個好用的東西。基本不需要設置，安上之後，用自己能登錄windows的帳號和密碼註冊就行了。雖然不是局域網上的機器，但也可以讓我的雙核CPU達到100%（平時開個Matlab什麼的都纔是50%，軟件優化真是關鍵啊）。
FOX算法有一些噁心的要求：輸入的矩陣必須是方陣，而且進程必須爲平方數，方陣必須能均勻劃分給每個進程。其實方陣的條件並不一定必須，只是會增加編程複雜性。
MPI中，雖然最精華的就6條語句，但僅用這6條，還是比較麻煩的。使用一些高級語句，能提高程序性能或是簡化代碼。
MPI_Bcast，這個是必須要提的。能把這樣的語句 if (task_id == root) { send to child_id } else { recieve from root} 僅用個bcast就替換了。
MPI_Isend/MPI_Irecv是異步發送和接收的語句，顯然比MPI_Send/MPI_Recv這樣的阻塞語句更可能提高性能。另外，還可以避免循環死鎖。當然，避免循環死鎖的方法還有使用奇偶法、MPI_Sendrecv等。
雖然，理解MPI_Cart_create等一系列的笛卡爾結構操作還是有些難度的，但卻是十分有用的進程邏輯功能劃分方法。特別在FOX算法中，每行是個通信子域，每列也是。這樣就會有n+n個子域（儘管有劃分重疊），每個域的進程都僅與域中的其它進程有信息交換，這樣僅用bcast一語就能傳完數據。
以下是代碼。比較簡單。就不說明了。
// MatrixMulFox.cpp// Jarod 2007.12.3#include "mpi.h"#include <algorithm>#include <fstream>#include <cmath>const int root_id = 0;const int max_procs_size = 16;int main(int argc,char *argv[])
{
double start_time, end_time, time; int procs_id, procs_size; MPI_Status status; MPI_Request reqSend, reqRecv; MPI_Init(&argc,&argv); start_time = MPI_Wtime(); MPI_Comm_size(MPI_COMM_WORLD,&procs_size); MPI_Comm_rank(MPI_COMM_WORLD,&procs_id); // 參數檢查 int N=0;
{
for (int i=1; i<argc; ++i ) { char * pos =strstr(argv[i], "-N="); if ( pos!=NULL) { sscanf(pos, "-N=%d", &N); break;
} } } const int procs_size_sqrt = floor(sqrt(static_cast<double>(procs_size))); const int n = N / procs_size_sqrt; const int n_sqr = n*n; if (procs_size<4 || procs_size> max_procs_size) { printf("The fox algorithm requires at least 4 processors and at most %d processors. ", max_procs_size); MPI_Finalize(); return 0; } if (procs_size_sqrt*procs_size_sqrt != procs_size ) { printf("The number of process must be a square. "); MPI_Finalize(); return 0; } if (N % procs_size_sqrt !=0) { printf("N mod procs_size_sqrt !=0 "); MPI_Finalize(); return 0; } //初始化矩陣 int * A = new int[n_sqr]; int * B = new int[n_sqr]; int * C = new int[n_sqr]; int * T = new int[n_sqr]; for (int i=0; i<n; ++i) for (int j=0; j<n; ++j) { A[i*n+j] = (i+j)*procs_id; B[i*n+j] = (i-j)*procs_id; C[i*n+j] = 0; } //輸出矩陣 printf("A on procs %d : ", procs_id); for (int i=0; i<n; ++i) { for (int j=0; j<n; ++j) { printf("%5d",A[i*n+j]); } printf(" "); } printf("B on procs %d : ", procs_id); for (int i=0; i<n; ++i) { for (int j=0; j<n; ++j) { printf("%5d",B[i*n+j]); } printf(" "); } // 劃分組, 建立子通信空間 MPI_Comm cart_all, cart_row, cart_col; int dims[2], periods[2]; int procs_cart_rank, procs_coords[2]; dims[0] = dims[1] = procs_size_sqrt; periods[0] = periods[1] = true; MPI_Cart_create(MPI_COMM_WORLD, 2, dims, periods, false, &cart_all); MPI_Comm_rank(cart_all, &procs_cart_rank); MPI_Cart_coords(cart_all, procs_cart_rank, 2, procs_coords); MPI_Comm_split(cart_all, procs_coords[0], procs_coords[1], &cart_row); MPI_Comm_split(cart_all, procs_coords[1], procs_coords[0], &cart_col); int rank_cart_row, rank_cart_col; MPI_Comm_rank(cart_row, & rank_cart_row); MPI_Comm_rank(cart_col, & rank_cart_col); // 計算並傳遞 for (int round = 0; round < procs_size_sqrt; ++ round) { MPI_Isend(B, n_sqr, MPI_INT, (procs_coords[0] - 1 + procs_size_sqrt) % procs_size_sqrt, 1, cart_col, &reqSend); int broader = (round + procs_coords[0]) % procs_size_sqrt; if (broader == procs_coords[1]) std::copy(A,A+n_sqr,T); MPI_Bcast(T, n_sqr, MPI_INT, broader , cart_row); for (int row=0; row<n; ++row) for (int col=0; col<n; ++col) for (int k=0; k<n; ++k) { C[row*n+col] += T[row*n+k]*B[k*n+col]; } MPI_Wait(&reqSend, &status); MPI_Recv(T, n_sqr, MPI_INT, (procs_coords[0] + 1) % procs_size_sqrt , 1, cart_col, &status); std::copy(T,T+n_sqr,B); } //輸出結果 printf("C on procs %d : ", procs_id); for (int i=0; i<n; ++i) { for (int j=0; j<n; ++j) { printf("%5d",C[i*n+j]); } printf(" "); } // 釋放 MPI_Comm_free(&cart_col); MPI_Comm_free(&cart_row); MPI_Comm_free(&cart_all); delete []A; delete []B; delete []C; delete []T; end_time = MPI_Wtime(); MPI_Finalize(); printf("task %d consumed %lf seconds ", procs_id, end_time-start_time); return 0;
}