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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
|
/*************************************************************************
* Copyright (c) 2015-2016, NVIDIA CORPORATION. All rights reserved.
*
* See LICENCE.txt for license information
************************************************************************/
#include <chrono>
#include <cstdio>
#include <cstdlib>
#include <string>
#include <vector>
#include "nccl.h"
#include "test_utilities.h"
template<typename T>
void RunTest(T** buff, const int N, const ncclDataType_t type, const int root,
ncclComm_t* const comms, const std::vector<int>& dList) {
// initialize data
int nDev = 0;
NCCLCHECK(ncclCommCount(comms[0], &nDev));
cudaStream_t* s = (cudaStream_t*)malloc(sizeof(cudaStream_t)*nDev);
T* buffer = (T*)malloc(N * sizeof(T));
T* result = (T*)malloc(N * sizeof(T));
memset(result, 0, N * sizeof(T));
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
CUDACHECK(cudaStreamCreate(s+i));
if (i == root) {
Randomize(buff[root], N, root);
CUDACHECK(cudaMemcpy(result, buff[root], N * sizeof(T),
cudaMemcpyDeviceToHost));
} else {
CUDACHECK(cudaMemset(buff[i], 0, N * sizeof(T)));
}
CUDACHECK(cudaDeviceSynchronize());
}
// warm up GPU
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
NCCLCHECK(ncclBcast((void*)buff[i], std::min(32 * 1024, N), type, root, comms[i], s[i]));
}
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
CUDACHECK(cudaStreamSynchronize(s[i]));
}
// for (int n = 1; n <= N; n = n << 1)
{
int n = N;
printf("%12i %12i %6s %4i", (int)(n * sizeof(T)), n,
TypeName(type).c_str(), root);
auto start = std::chrono::high_resolution_clock::now();
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
NCCLCHECK(ncclBcast((void*)buff[i], n, type, root, comms[i], s[i]));
}
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
CUDACHECK(cudaStreamSynchronize(s[i]));
}
auto stop = std::chrono::high_resolution_clock::now();
double elapsedSec =
std::chrono::duration_cast<std::chrono::duration<double>>(
stop - start).count();
double algbw = (double)(n * sizeof(T)) / 1.0E9 / elapsedSec;
double busbw = algbw;
double maxDelta = 0.0;
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
double tmpDelta = CheckDelta<T>(buff[i], result, n);
maxDelta = std::max(tmpDelta, maxDelta);
}
printf(" %7.3f %5.2f %5.2f %7.0le\n", elapsedSec * 1.0E3, algbw, busbw,
maxDelta);
}
for(int i=0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
CUDACHECK(cudaStreamDestroy(s[i]));
}
free(s);
free(buffer);
free(result);
}
template<typename T>
void RunTests(const int N, const ncclDataType_t type, ncclComm_t* const comms,
const std::vector<int>& dList) {
int nDev = 0;
NCCLCHECK(ncclCommCount(comms[0], &nDev));
T** buff = (T**)malloc(nDev * sizeof(T*));
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
CUDACHECK(cudaMalloc(buff + i, N * sizeof(T)));
}
//for (int root = 1; root < 2; ++root) {
for (int root = 0; root < nDev; ++root) {
RunTest<T>(buff, N, type, root, comms, dList);
}
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(dList[i]));
CUDACHECK(cudaFree(buff[i]));
}
free(buff);
}
void usage() {
printf("Tests nccl Broadcast with user supplied arguments.\n"
" Usage: broadcast_test <data size in bytes> [number of GPUs] "
"[GPU 0] [GPU 1] ...\n\n");
}
int main(int argc, char* argv[]) {
int nVis = 0;
CUDACHECK(cudaGetDeviceCount(&nVis));
unsigned long long N = 0;
if (argc > 1) {
int t = sscanf(argv[1], "%llu", &N);
if (t == 0) {
printf("Error: %s is not an integer!\n\n", argv[1]);
usage();
exit(EXIT_FAILURE);
}
} else {
printf("Error: must specify at least data size in bytes!\n\n");
usage();
exit(EXIT_FAILURE);
}
int nDev = nVis;
if (argc > 2) {
int t = sscanf(argv[2], "%d", &nDev);
if (t == 0) {
printf("Error: %s is not an integer!\n\n", argv[1]);
usage();
exit(EXIT_FAILURE);
}
}
std::vector<int> dList(nDev);
for (int i = 0; i < nDev; ++i)
dList[i] = i % nVis;
if (argc > 3) {
if (argc - 3 != nDev) {
printf("Error: insufficient number of GPUs in list\n\n");
usage();
exit(EXIT_FAILURE);
}
for (int i = 0; i < nDev; ++i) {
int t = sscanf(argv[3 + i], "%d", dList.data() + i);
if (t == 0) {
printf("Error: %s is not an integer!\n\n", argv[2 + i]);
usage();
exit(EXIT_FAILURE);
}
}
}
ncclComm_t* comms = (ncclComm_t*)malloc(sizeof(ncclComm_t)*nDev);;
NCCLCHECK(ncclCommInitAll(comms, nDev, dList.data()));
printf("# Using devices\n");
for (int g = 0; g < nDev; ++g) {
int cudaDev;
int rank;
cudaDeviceProp prop;
NCCLCHECK(ncclCommCuDevice(comms[g], &cudaDev));
NCCLCHECK(ncclCommUserRank(comms[g], &rank));
CUDACHECK(cudaGetDeviceProperties(&prop, cudaDev));
printf("# Rank %2d uses device %2d [0x%02x] %s\n", rank, cudaDev,
prop.pciBusID, prop.name);
}
printf("\n");
printf("# %10s %12s %6s %4s %7s %5s %5s %7s\n",
"bytes", "N", "type", "root", "time", "algbw", "busbw", "delta");
RunTests<char>(N / sizeof(char), ncclChar, comms, dList);
RunTests<int>(N / sizeof(int), ncclInt, comms, dList);
#ifdef CUDA_HAS_HALF
RunTests<half>(N / sizeof(half), ncclHalf, comms, dList);
#endif
RunTests<float>(N / sizeof(float), ncclFloat, comms, dList);
RunTests<double>(N / sizeof(double), ncclDouble, comms, dList);
RunTests<long long>(N / sizeof(long long), ncclInt64, comms, dList);
RunTests<unsigned long long>(N / sizeof(unsigned long long), ncclUint64, comms, dList);
printf("\n");
for(int i = 0; i < nDev; ++i)
ncclCommDestroy(comms[i]);
free(comms);
exit(EXIT_SUCCESS);
}
|
