1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
| #include <chrono>
#include <iostream>
#include <mpi.h>
#include <time.h>
#include <cstring>
#include <cmath>
#include <algorithm>
#define EPS 1e-5
namespace ch = std::chrono;
void Ring_Allreduce(void* sendbuf, void* recvbuf, int n, MPI_Comm comm, int comm_sz, int my_rank)
{
int chunk = (n + comm_sz - 1) / comm_sz;
int last_chunk = n - (comm_sz-1)*chunk;
int total_bytes = n * sizeof(float);
float* sendbuf_float = (float*)sendbuf;
float* recvbuf_float = (float*)recvbuf;
float* tmpbuf_float = (float*)malloc(total_bytes);
memcpy(recvbuf, sendbuf, total_bytes);
for (int i = 0; i < comm_sz - 1; ++i){
MPI_Request send_req,recv_req;
if ((my_rank - i + comm_sz) % comm_sz == comm_sz - 1){
MPI_Isend(recvbuf_float + (my_rank - i + comm_sz) % comm_sz *chunk, last_chunk, MPI_FLOAT, (my_rank + 1) % comm_sz, 0, comm,&send_req);
MPI_Irecv(tmpbuf_float+(my_rank - 1 +comm_sz - i) % comm_sz *chunk, chunk, MPI_FLOAT, (my_rank - 1 + comm_sz) % comm_sz, 0, comm,&recv_req);
}else if ((my_rank -1 + comm_sz - i) % comm_sz == comm_sz -1){
MPI_Isend(recvbuf_float + (my_rank - i + comm_sz) % comm_sz *chunk, chunk, MPI_FLOAT, (my_rank + 1) % comm_sz, 0, comm,&send_req);
MPI_Irecv(tmpbuf_float+(my_rank - 1 +comm_sz - i) % comm_sz *chunk, last_chunk, MPI_FLOAT, (my_rank - 1 + comm_sz) % comm_sz, 0, comm,&recv_req);
}else{
MPI_Isend(recvbuf_float + (my_rank - i + comm_sz) % comm_sz *chunk , chunk, MPI_FLOAT, (my_rank + 1) % comm_sz, 0, comm,&send_req);
MPI_Irecv(tmpbuf_float+(my_rank - 1 +comm_sz - i) % comm_sz *chunk , chunk, MPI_FLOAT, (my_rank - 1 + comm_sz) % comm_sz, 0, comm,&recv_req);
}
MPI_Wait(&send_req, MPI_STATUS_IGNORE);
MPI_Wait(&recv_req, MPI_STATUS_IGNORE);
int special_chunk = chunk;
if ((my_rank -1 + comm_sz - i) % comm_sz == comm_sz -1){
special_chunk = last_chunk;
}
for (int j = 0; j < special_chunk; ++j){
recvbuf_float[j+(my_rank - 1 +comm_sz - i) % comm_sz*chunk] += tmpbuf_float[j + (my_rank - 1 -i + comm_sz) % comm_sz*chunk];
}
}
for (int i = 0; i < comm_sz-1; ++i){
MPI_Request send_req,recv_req;
int source_chunk = (my_rank + 1 - i + comm_sz) % comm_sz;
int dest_chunk = (my_rank - i + comm_sz) % comm_sz;
if (source_chunk == comm_sz - 1){
MPI_Isend(recvbuf_float + source_chunk*chunk, last_chunk, MPI_FLOAT, (my_rank + 1) % comm_sz, 0, comm,&send_req);
MPI_Irecv(recvbuf_float + dest_chunk*chunk, chunk, MPI_FLOAT, (my_rank - 1 + comm_sz) % comm_sz, 0, comm,&recv_req);
}else if (dest_chunk == comm_sz - 1){
MPI_Isend(recvbuf_float + source_chunk*chunk, chunk, MPI_FLOAT, (my_rank + 1) % comm_sz, 0, comm,&send_req);
MPI_Irecv(recvbuf_float + dest_chunk*chunk,last_chunk, MPI_FLOAT, (my_rank - 1 + comm_sz) % comm_sz, 0, comm,&recv_req);
}else{
MPI_Isend(recvbuf_float + source_chunk*chunk, chunk, MPI_FLOAT, (my_rank + 1) % comm_sz, 0, comm,&send_req);
MPI_Irecv(recvbuf_float + dest_chunk*chunk, chunk, MPI_FLOAT, (my_rank - 1 + comm_sz) % comm_sz, 0, comm,&recv_req);
}
MPI_Wait(&send_req, MPI_STATUS_IGNORE);
MPI_Wait(&recv_req, MPI_STATUS_IGNORE);
}
free(tmpbuf_float);
}
void Naive_Allreduce(void* sendbuf, void* recvbuf, int n, MPI_Comm comm, int comm_sz, int my_rank)
{
MPI_Reduce(sendbuf, recvbuf, n, MPI_FLOAT, MPI_SUM, 0, comm);
MPI_Bcast(recvbuf, n, MPI_FLOAT, 0, comm);
}
int main(int argc, char *argv[])
{
int ITER = atoi(argv[1]);
int n = atoi(argv[2]);
float* mpi_sendbuf = new float[n];
float* mpi_recvbuf = new float[n];
float* naive_sendbuf = new float[n];
float* naive_recvbuf = new float[n];
float* ring_sendbuf = new float[n];
float* ring_recvbuf = new float[n];
MPI_Init(nullptr, nullptr);
int comm_sz;
int my_rank;
MPI_Comm_size(MPI_COMM_WORLD, &comm_sz);
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
srand(time(NULL) + my_rank);
for (int i = 0; i < n; ++i)
mpi_sendbuf[i] = static_cast <float> (rand()) / static_cast <float> (RAND_MAX);
memcpy(naive_sendbuf, mpi_sendbuf, n * sizeof(float));
memcpy(ring_sendbuf, mpi_sendbuf, n * sizeof(float));
MPI_Allreduce(mpi_sendbuf, mpi_recvbuf, n, MPI_FLOAT, MPI_SUM, MPI_COMM_WORLD);
Naive_Allreduce(naive_sendbuf, naive_recvbuf, n, MPI_COMM_WORLD, comm_sz, my_rank);
Ring_Allreduce(ring_sendbuf, ring_recvbuf, n, MPI_COMM_WORLD, comm_sz, my_rank);
bool correct = true;
for (int i = 0; i < n; ++i)
if (abs(mpi_recvbuf[i] - ring_recvbuf[i]) > EPS)
{
correct = false;
break;
}
if (correct)
{
auto beg = ch::high_resolution_clock::now();
for (int iter = 0; iter < ITER; ++iter)
MPI_Allreduce(mpi_sendbuf, mpi_recvbuf, n, MPI_FLOAT, MPI_SUM, MPI_COMM_WORLD);
auto end = ch::high_resolution_clock::now();
double mpi_dur = ch::duration_cast<ch::duration<double>>(end - beg).count() * 1000;
beg = ch::high_resolution_clock::now();
for (int iter = 0; iter < ITER; ++iter)
Naive_Allreduce(naive_sendbuf, naive_recvbuf, n, MPI_COMM_WORLD, comm_sz, my_rank);
end = ch::high_resolution_clock::now();
double naive_dur = ch::duration_cast<ch::duration<double>>(end - beg).count() * 1000;
beg = ch::high_resolution_clock::now();
for (int iter = 0; iter < ITER; ++iter)
Ring_Allreduce(ring_sendbuf, ring_recvbuf, n, MPI_COMM_WORLD, comm_sz, my_rank);
end = ch::high_resolution_clock::now();
double ring_dur = ch::duration_cast<ch::duration<double>>(end - beg).count() * 1000;
if (my_rank == 0)
{
std::cout << "Correct." << std::endl;
std::cout << "MPI_Allreduce: " << mpi_dur << " ms." << std::endl;
std::cout << "Naive_Allreduce: " << naive_dur << " ms." << std::endl;
std::cout << "Ring_Allreduce: " << ring_dur << " ms." << std::endl;
}
}
else
if (my_rank == 0)
std::cout << "Wrong!" << std::endl;
delete[] mpi_sendbuf;
delete[] mpi_recvbuf;
delete[] naive_sendbuf;
delete[] naive_recvbuf;
delete[] ring_sendbuf;
delete[] ring_recvbuf;
MPI_Finalize();
return 0;
}
|
