/
Checksum_implementation.c
358 lines (322 loc) · 13.9 KB
/
Checksum_implementation.c
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
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
#include "main_funcs.h"
extern "C" void openblas_set_num_threads(int);
using namespace std;
//Checksum implementation. 3 nodes for data
int main (int argc, char *argv[])
{
if(argc<4)
{
printf("ERROR, argc is %d\n",argc);
return 0;
}
time_to_run = atoi(argv[1]);
query_size=atoi(argv[2]);
data_size=atoi(argv[3]);
//set the number of nodes and number of threads per node on the main_funcs.h accordingly
omp_set_num_threads(nnodes);
openblas_set_num_threads(1);
if((data_size%total_threads!=0) || (data_size%total_threads_chksum!=0))
{
printf("Wrong dbsize. Choose dbsize divisible by %d and %d\n",total_threads,total_threads_chksum);
return 0;
}
data_size_per_thread_chksum = data_size/total_threads_chksum;
float* Query_row = (float*)malloc(data_dim*query_size*sizeof(float));
float** Data_perthread = (float**)malloc(total_threads*sizeof(float*));
float** Result_perthread = (float**)malloc(total_threads*sizeof(float*));
float** top_three_each_node = allocarray(query_size,12);
startwtime=0.0;
average_time=0;
for(int i=0; i< total_threads;i++){ //16 total threads
Data_perthread[i] = (float*)malloc(data_size_per_thread_chksum*data_dim*sizeof(float));
Result_perthread[i] = (float*)malloc(data_size_per_thread_chksum*query_size*sizeof(float));
}
__m128 load_data,load_data1,add_data;
/*if(data_size_per_thread_chksum>9344)
{
printf("Please choose a smaller dbsize\n");
//because 9344 is the max we have;
return 0;
}*/
char Queryfile[300],*pathtofile; //change this section to suit naming convention of Queryfile.
pathtofile = "***"; //absolute path_to_query bin file eg. ./Image/Queryfile288.bin
strcpy(Queryfile,pathtofile);
/*
char str_query_size[10];
snprintf(str_query_size,10,"%d",query_size);
strcat(Queryfile,str_query_size);
strcat(Queryfile,ext);*/
char *ext;
char* database_path = ""; //absolute path_to_database bin file eg. ./Image/ConventionalDB82944.bin
char str_threads_used[10]; //modify this based on your DB naming convention
snprintf(str_threads_used,10,"%d",total_threads);
ext= ".bin";
startwtime=0.0;
average_time=0;
char *ext;
char* database_path = ""; //absolute path_to_database bin file eg. ./Image/ChecksumDB82944.bin
char str_threads_used[10]; //modify this based on your DB naming convention
snprintf(str_threads_used,10,"%d",total_threads);
ext= ".bin";
//Ensure Database for checksum node contains checksum data
MPI_Status status;
/***** Initializations *****/
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numtasks);
MPI_Comm_rank(MPI_COMM_WORLD,&taskid);
printf ("MPI task %d has started...\n", taskid);
/***** Master task only ******/
if (taskid == MASTER){ //we use the xxx12.bin files as they contain the checksum
printf("Checksum with query_size= %d running %d times\n",query_size,time_to_run);
printf("Actual nodes used: %d\n",numtasks-1);
Query_row= ReadFile(Queryfile,query_size);
for(int i=0; i<total_threads; i++) {
char threadname[10]; //modify this section based on your DB naming convention
snprintf(threadname,10,"%d",i);
char myfilename[300];
strcpy(myfilename,database_path);
strcat(myfilename,threadname);
strcat(myfilename,str_threads_used);
strcat(myfilename,extension);
Data_perthread[i] = ReadFile(myfilename,data_size_per_thread_chksum);
}
for(int i=1;i<numtasks;i++)
{
//here
printf("sending....\n");
for (int j=0;j<nthreads_per_node;j++)
{
MPI_Send(&Data_perthread[(i-1)*nthreads_per_node][0],data_size_per_thread_chksum*data_dim,MPI_FLOAT,i,j+1,MPI_COMM_WORLD);
printf("sent %d to node %d\n",j+1,i);
}
}
float** All_top_three = allocarray(query_size, 3*total_threads_chksum) ;
MPI_Datatype mysubarray;
int starts[2] = {0,0};
int subsizes[2] = {query_size,3*nnodes};
int bigsizes[2] = {query_size,3*total_threads_chksum};
MPI_Type_create_subarray(2,bigsizes, subsizes, starts, MPI_ORDER_C, MPI_INT, &mysubarray);
MPI_Type_commit(&mysubarray);
for(int ij=0;ij<time_to_run;ij++)
{
startwtime= MPI_Wtime();
for(int i=1;i<numtasks;i++)
{
MPI_Send(&Query_row[0],query_size*data_dim,MPI_FLOAT,i,nthreads_per_node+1,MPI_COMM_WORLD);
}
//receive from all but taskid = 1
//final sorting
for(int i=2;i<(numtasks-1);i++)
{
MPI_Recv(&(All_top_three[0][(i-1)*3*nnodes]), 1, mysubarray, i, nthreads_per_node+21, MPI_COMM_WORLD,&status);
}
MPI_Recv(&(All_top_three[0][0]), 1, mysubarray, (numtasks-1), nthreads_per_node+21, MPI_COMM_WORLD,&status);
//final sorting, we just get the master to do this joor
#pragma omp parallel for
for(int i=0;i<query_size;i++)
{
sort(All_top_three[i],All_top_three[i]+3*total_threads_chksum,greaterthanfunction);
}
endwtime= MPI_Wtime();
average_time += (endwtime-startwtime);
printf("Time to complete: %f\n",endwtime-startwtime);
}
printf("\nAverage Time to complete Checksum: %f secs\n\n",(average_time/time_to_run));
MPI_Type_free(&mysubarray);
} /* end of master section */
/***** Non-master tasks only *****/
else if (taskid > MASTER ) {
for(int i=0; i<nthreads_per_node;i++)
{
MPI_Recv(&Data_perthread[(taskid-1)*nthreads_per_node][0],data_size_per_thread_chksum*data_dim,MPI_FLOAT,0,i+1,MPI_COMM_WORLD,&status);
}
for(int ij=0;ij<time_to_run;ij++)
{
MPI_Recv(&Query_row[0],query_size*data_dim,MPI_FLOAT,0,nthreads_per_node+1,MPI_COMM_WORLD,&status);
#pragma omp parallel
{
int id= omp_get_thread_num();
cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,query_size,data_size_per_thread_chksum, data_dim, 1.0,Query_row,data_dim,Data_perthread[(taskid-1)*nthreads_per_node +id],data_dim,0.0,Result_perthread[(taskid-1)*nthreads_per_node +id],data_size_per_thread_chksum);
}
//we assume taskid 1 failed
//this experiment is for 4 nodes only. Please design the Checksum implentation to suport more than this.
if(taskid==2)
{
//This node has data for threads 4 and 5 already as he computed them
MPI_Recv(&Result_perthread[8][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,3,10,MPI_COMM_WORLD,&status);
MPI_Recv(&Result_perthread[12][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,4,11,MPI_COMM_WORLD,&status);
MPI_Recv(&Result_perthread[9][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,3,13,MPI_COMM_WORLD,&status);
MPI_Recv(&Result_perthread[13][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,4,14,MPI_COMM_WORLD,&status);
MPI_Send(&Result_perthread[6][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,3,15,MPI_COMM_WORLD);
MPI_Send(&Result_perthread[7][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,4,18,MPI_COMM_WORLD);
//recover
#pragma omp parallel private(load_data,load_data1,add_data)
{
#pragma omp for
for(int i=0;i<query_size*data_size_per_thread_chksum/4;i++)
{
int j = i*4;
load_data = _mm_load_ps(Result_perthread[4]+j);
load_data1 = _mm_load_ps(Result_perthread[8]+j);
add_data = _mm_load_ps(Result_perthread[12]+j);
add_data=(_mm_sub_ps(add_data, _mm_add_ps(load_data,load_data1)));
_mm_store_ps(Result_perthread[0]+j, add_data);
}
#pragma omp for
for(int i=0;i<query_size*data_size_per_thread_chksum/4;i++)
{
int j = i*4;
load_data = _mm_load_ps(Result_perthread[5]+j);
load_data1 = _mm_load_ps(Result_perthread[9]+j);
add_data = _mm_load_ps(Result_perthread[13]+j);
add_data=(_mm_sub_ps(add_data, _mm_add_ps(load_data,load_data1)));
_mm_store_ps(Result_perthread[1]+j, add_data);
}
}
//Send the recovered to 4 to sort
MPI_Send(&Result_perthread[0][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,4,21,MPI_COMM_WORLD);
MPI_Send(&Result_perthread[1][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,4,22,MPI_COMM_WORLD);
#pragma omp parallel private (values_to_retain,value_in_mat)
{
int id = omp_get_thread_num();
for (int loc=0;loc<query_size;loc++)
{
values_to_retain.clear();
values_to_retain.resize(data_size_per_thread_chksum);
int counts=0;
for(int ind=0;ind<data_size_per_thread_chksum;ind++)
{
value_in_mat = Result_perthread[(taskid-1)*4+id][loc*data_size_per_thread_chksum + ind];
if(value_in_mat>threshold)
{
values_to_retain[counts]= value_in_mat;
counts++;
}
}
sort(values_to_retain.begin(),values_to_retain.end(),greaterthanfunction);
//Here we are putting 3 top images per core in a
top_three_each_node[loc][id*3] =values_to_retain[0];
top_three_each_node[loc][id*3 +1] = values_to_retain[1];
top_three_each_node[loc][id*3+2] = values_to_retain[2];
}
}
MPI_Datatype mysubarray;
int starts[2] = {0,0};
int subsizes[2] = {query_size,12};
MPI_Type_create_subarray(2, subsizes, subsizes, starts, MPI_ORDER_C, MPI_INT, &mysubarray);
MPI_Type_commit(&mysubarray);
MPI_Send(&(top_three_each_node[0][0]), 1, mysubarray, 0, 25, MPI_COMM_WORLD);
MPI_Type_free(&mysubarray);
}
if(taskid==3)
{
MPI_Send(&Result_perthread[8][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,2,10,MPI_COMM_WORLD);
MPI_Send(&Result_perthread[9][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,2,13,MPI_COMM_WORLD);
MPI_Recv(&Result_perthread[6][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,2,15,MPI_COMM_WORLD,&status);
MPI_Recv(&Result_perthread[14][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,4,17,MPI_COMM_WORLD,&status);
MPI_Send(&Result_perthread[11][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,4,19,MPI_COMM_WORLD);
#pragma omp parallel private(load_data,load_data1,add_data)
{
#pragma omp for
for(int i=0;i<query_size*data_size_per_thread_chksum/4;i++)
{
int j = i*4;
load_data = _mm_load_ps(Result_perthread[6]+j);
load_data1 = _mm_load_ps(Result_perthread[10]+j);
add_data = _mm_load_ps(Result_perthread[14]+j);
add_data=(_mm_sub_ps(add_data, _mm_add_ps(load_data,load_data1)));
_mm_store_ps(Result_perthread[2]+j, add_data);
}
}
MPI_Send(&Result_perthread[2][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,4,23,MPI_COMM_WORLD);
#pragma omp parallel private (values_to_retain,value_in_mat)
{
int id = omp_get_thread_num();
for (int loc=0;loc<query_size;loc++)
{
values_to_retain.clear();
values_to_retain.resize(data_size_per_thread_chksum);
int counts=0;
for(int ind=0;ind<data_size_per_thread_chksum;ind++)
{
value_in_mat = Result_perthread[(taskid-1)*4+id][loc*data_size_per_thread_chksum + ind];
if(value_in_mat>threshold)
{
values_to_retain[counts]= value_in_mat;
counts++;
}
}
sort(values_to_retain.begin(),values_to_retain.end(),greaterthanfunction);
top_three_each_node[loc][id*3] =values_to_retain[0];
top_three_each_node[loc][id*3 +1] = values_to_retain[1];
top_three_each_node[loc][id*3+2] = values_to_retain[2];
}
}
MPI_Datatype mysubarray;
int starts[2] = {0,0};
int subsizes[2] = {query_size,12};
MPI_Type_create_subarray(2, subsizes, subsizes, starts, MPI_ORDER_C, MPI_INT, &mysubarray);
MPI_Type_commit(&mysubarray);
MPI_Send(&(top_three_each_node[0][0]), 1, mysubarray, 0, 25, MPI_COMM_WORLD);
MPI_Type_free(&mysubarray);
}
if(taskid==4)
{
MPI_Send(&Result_perthread[12][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,2,11,MPI_COMM_WORLD);
MPI_Send(&Result_perthread[13][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,2,14,MPI_COMM_WORLD);
MPI_Send(&Result_perthread[14][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,3,17,MPI_COMM_WORLD);
MPI_Recv(&Result_perthread[7][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,2,18,MPI_COMM_WORLD,&status);
MPI_Recv(&Result_perthread[11][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,3,19,MPI_COMM_WORLD,&status);
#pragma omp parallel private(load_data,load_data1,add_data,i,j)
{
#pragma omp for
for(int i=0;i<query_size*data_size_per_thread_chksum/4;i++)
{
int j = i*4;
load_data = _mm_load_ps(Result_perthread[7]+j);
load_data1 = _mm_load_ps(Result_perthread[11]+j);
add_data = _mm_load_ps(Result_perthread[15]+j);
add_data=(_mm_sub_ps(add_data, _mm_add_ps(load_data,load_data1)));
_mm_store_ps(Result_perthread[3]+j, add_data);
}
}
MPI_Recv(&Result_perthread[2][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,3,23,MPI_COMM_WORLD,&status);
MPI_Recv(&Result_perthread[0][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,2,21,MPI_COMM_WORLD,&status);
MPI_Recv(&Result_perthread[1][0],query_size*data_size_per_thread_chksum,MPI_FLOAT,2,22,MPI_COMM_WORLD,&status);
#pragma omp parallel private (values_to_retain,value_in_mat)
{
int id = omp_get_thread_num();
for (int loc=0;loc<query_size;loc++)
{
values_to_retain.clear();
values_to_retain.resize(data_size_per_thread_chksum);
int counts=0;
for(int ind=0;ind<data_size_per_thread_chksum;ind++)
{
value_in_mat = Result_perthread[id][loc*data_size_per_thread_chksum + ind];
if(value_in_mat>threshold)
{
values_to_retain[counts]= value_in_mat;
counts++;
}
}
sort(values_to_retain.begin(),values_to_retain.end(),greaterthanfunction);
//Here we are putting 3 top images per core in a
top_three_each_node[loc][id*3] =values_to_retain[0];
top_three_each_node[loc][id*3 +1] = values_to_retain[1];
top_three_each_node[loc][id*3+2] = values_to_retain[2];
}
}
MPI_Datatype mysubarray;
int starts[2] = {0,0};
int subsizes[2] = {query_size,12};
MPI_Type_create_subarray(2, subsizes, subsizes, starts, MPI_ORDER_C, MPI_INT, &mysubarray);
MPI_Type_commit(&mysubarray);
MPI_Send(&(top_three_each_node[0][0]), 1, mysubarray, 0, 25, MPI_COMM_WORLD);
MPI_Type_free(&mysubarray);
}
}
}
MPI_Finalize();
return 0;
}