| 1 |
#include <sys/time.h> |
| 2 |
#include <string.h> |
| 3 |
#include <stdio.h> |
| 4 |
#include <stdlib.h> |
| 5 |
#include <math.h> |
| 6 |
|
| 7 |
#ifdef IS_MPI |
| 8 |
#include <mpi.h> |
| 9 |
|
| 10 |
#include "mpiSimulation.hpp" |
| 11 |
#endif //is_mpi |
| 12 |
|
| 13 |
#include "simError.h" |
| 14 |
#include "mdProfile.hpp" |
| 15 |
|
| 16 |
namespace mdProfileSpace { |
| 17 |
|
| 18 |
class ProfileString{ |
| 19 |
public: |
| 20 |
char myName[MAX_PROFILE_NAMELENGTH]; |
| 21 |
}; |
| 22 |
|
| 23 |
ProfileString theNames[N_PROFILES]; |
| 24 |
|
| 25 |
struct timeval startTime[N_PROFILES]; |
| 26 |
struct timeval endTime[N_PROFILES]; |
| 27 |
|
| 28 |
double accumTime[N_PROFILES]; |
| 29 |
|
| 30 |
#ifdef IS_MPI |
| 31 |
double globalTime[N_PROFILES]; |
| 32 |
#endif //is_mpi |
| 33 |
|
| 34 |
} |
| 35 |
|
| 36 |
|
| 37 |
using namespace mdProfileSpace; |
| 38 |
|
| 39 |
|
| 40 |
void initProfile( void ){ |
| 41 |
|
| 42 |
int i; |
| 43 |
|
| 44 |
for( i=0;i<N_PROFILES;i++ ){ |
| 45 |
|
| 46 |
accumTime[i] = 0.0; |
| 47 |
|
| 48 |
#ifdef IS_MPI |
| 49 |
globalTime[i] = 0.0; |
| 50 |
#endif //is_mpi |
| 51 |
} |
| 52 |
|
| 53 |
strncpy( theNames[pro1].myName, "BASS Parse", MAX_PROFILE_NAMELENGTH ); |
| 54 |
strncpy( theNames[pro2].myName, "SimSetup->CreateSim", MAX_PROFILE_NAMELENGTH ); |
| 55 |
strncpy( theNames[pro3].myName, "Integrate the Sim", MAX_PROFILE_NAMELENGTH ); |
| 56 |
strncpy( theNames[pro4].myName, "unamed4", MAX_PROFILE_NAMELENGTH ); |
| 57 |
strncpy( theNames[pro5].myName, "unamed5", MAX_PROFILE_NAMELENGTH ); |
| 58 |
strncpy( theNames[pro6].myName, "unamed6", MAX_PROFILE_NAMELENGTH ); |
| 59 |
strncpy( theNames[pro7].myName, "unamed7", MAX_PROFILE_NAMELENGTH ); |
| 60 |
strncpy( theNames[pro8].myName, "unamed8", MAX_PROFILE_NAMELENGTH ); |
| 61 |
} |
| 62 |
|
| 63 |
|
| 64 |
void startProfile( proNames theProfile ){ |
| 65 |
struct timezone tz; |
| 66 |
|
| 67 |
gettimeofday( &startTime[theProfile], &tz ); |
| 68 |
} |
| 69 |
|
| 70 |
void endProfile( proNames theProfile ){ |
| 71 |
struct timezone tz; |
| 72 |
double startVal, endVal; |
| 73 |
|
| 74 |
gettimeofday( &endTime[theProfile], &tz ); |
| 75 |
|
| 76 |
startVal = (double)startTime[theProfile].tv_sec |
| 77 |
+ (double)startTime[theProfile].tv_usec / 1000000.0; |
| 78 |
|
| 79 |
endVal = (double)endTime[theProfile].tv_sec |
| 80 |
+ (double)endTime[theProfile].tv_usec / 1000000.0; |
| 81 |
|
| 82 |
accumTime[theProfile] += endVal - startVal; |
| 83 |
} |
| 84 |
|
| 85 |
|
| 86 |
void writeProfiles( void ){ |
| 87 |
|
| 88 |
int i; |
| 89 |
double totalTime; |
| 90 |
double percentTime[N_PROFILES]; |
| 91 |
int days, hours, minutes, secs, msecs; |
| 92 |
double donkey; |
| 93 |
|
| 94 |
#ifdef IS_MPI |
| 95 |
int j; |
| 96 |
|
| 97 |
MPI_Status istatus; |
| 98 |
|
| 99 |
double nodeTime; |
| 100 |
double nodeAccum[N_PROFILES]; |
| 101 |
double nodePercent[N_PROFILES]; |
| 102 |
|
| 103 |
double globalTime; |
| 104 |
double globalAccum[N_PROFILES]; |
| 105 |
double globalPercent[N_PROFILES]; |
| 106 |
#endif // is_mpi |
| 107 |
|
| 108 |
|
| 109 |
#ifndef IS_MPI // single processor version |
| 110 |
|
| 111 |
totalTime = 0.0; |
| 112 |
for(i=0;i<N_PROFILES;i++) |
| 113 |
totalTime += accumTime[i]; |
| 114 |
|
| 115 |
for(i=0;i<N_PROFILES;i++) |
| 116 |
percentTime[i] = accumTime[i] / totalTime; |
| 117 |
|
| 118 |
fprintf(stdout, |
| 119 |
" Time Spent Percent Time Name\n" |
| 120 |
"-------------- ---------------- -----------------------------------------\n" |
| 121 |
); |
| 122 |
|
| 123 |
for(i=0;i<N_PROFILES;i++){ |
| 124 |
fprintf(stdout, |
| 125 |
" %12G %14G %40s\n", |
| 126 |
accumTime[i], |
| 127 |
percentTime[i], |
| 128 |
theNames[i].myName ); |
| 129 |
} |
| 130 |
|
| 131 |
days = (int)floor( totalTime / 86400 ); |
| 132 |
donkey = totalTime - 86400 * days; |
| 133 |
|
| 134 |
hours = (int)floor( donkey / 3600 ); |
| 135 |
donkey -= hours * 3600; |
| 136 |
|
| 137 |
minutes = (int)floor( donkey / 60 ); |
| 138 |
donkey -= minutes * 60; |
| 139 |
|
| 140 |
secs = (int)donkey; |
| 141 |
msecs = (int)( (donkey - secs) * 1000 ); |
| 142 |
|
| 143 |
fprintf( stdout, |
| 144 |
"----------------------------------------------------------------------------\n" |
| 145 |
" Total Time = %03d:%02d:%02d:%02d.%03d ( %G sec )\n", |
| 146 |
days, |
| 147 |
hours, |
| 148 |
minutes, |
| 149 |
secs, |
| 150 |
msecs, |
| 151 |
totalTime ); |
| 152 |
|
| 153 |
#else // the parrallel version |
| 154 |
|
| 155 |
if( worldRank == 0 ){ |
| 156 |
|
| 157 |
double *nodeTots = new double[mpiSim->getNumberProcessors()]; |
| 158 |
double *nodePercentTots = new double[mpiSim->getNumberProcessors()]; |
| 159 |
|
| 160 |
totalTime = 0.0; |
| 161 |
for(i=0;i<N_PROFILES;i++) |
| 162 |
totalTime += accumTime[i]; |
| 163 |
|
| 164 |
for(i=0;i<N_PROFILES;i++) |
| 165 |
percentTime[i] = accumTime[i] / totalTime; |
| 166 |
|
| 167 |
fprintf(stdout, |
| 168 |
"\n" |
| 169 |
"----------------------------------------------------------------------------\n" |
| 170 |
" Output from Node %d: \n" |
| 171 |
"\n" |
| 172 |
" Time Spent Percent Time Name\n" |
| 173 |
"-------------- ---------------- -----------------------------------------\n", |
| 174 |
worldRank); |
| 175 |
|
| 176 |
for(i=0;i<N_PROFILES;i++){ |
| 177 |
fprintf(stdout, |
| 178 |
" %12G %14G %40s\n", |
| 179 |
accumTime[i], |
| 180 |
percentTime[i], |
| 181 |
theNames[i].myName ); |
| 182 |
} |
| 183 |
|
| 184 |
days = (int)floor( totalTime / 86400 ); |
| 185 |
donkey = totalTime - 86400 * days; |
| 186 |
|
| 187 |
hours = (int)floor( donkey / 3600 ); |
| 188 |
donkey -= hours * 3600; |
| 189 |
|
| 190 |
minutes = (int)floor( donkey / 60 ); |
| 191 |
donkey -= minutes * 60; |
| 192 |
|
| 193 |
secs = (int)donkey; |
| 194 |
msecs = (int)( (donkey - secs) * 1000 ); |
| 195 |
|
| 196 |
fprintf( stdout, |
| 197 |
"----------------------------------------------------------------------------\n" |
| 198 |
" Total Time = %03d:%02d:%02d:%02d.%03d ( %G sec )\n", |
| 199 |
days, |
| 200 |
hours, |
| 201 |
minutes, |
| 202 |
secs, |
| 203 |
msecs, |
| 204 |
totalTime ); |
| 205 |
|
| 206 |
// now the rest of the nodes |
| 207 |
|
| 208 |
nodeTots[0] = totalTime; |
| 209 |
|
| 210 |
globalTime = totalTime; |
| 211 |
for(i=0;i<N_PROFILES;i++) |
| 212 |
globalAccum[i] = accumTime[i]; |
| 213 |
|
| 214 |
|
| 215 |
for(j=1;j<mpiSim->getNumberProcessors();j++){ |
| 216 |
|
| 217 |
nodeTime = 0.0; |
| 218 |
|
| 219 |
MPI_Recv(nodeAccum, N_PROFILES, MPI_DOUBLE, j, |
| 220 |
1, MPI_COMM_WORLD, &istatus ); |
| 221 |
|
| 222 |
for(i=0;i<N_PROFILES;i++){ |
| 223 |
nodeTime += nodeAccum[i]; |
| 224 |
} |
| 225 |
|
| 226 |
for(i=0;i<N_PROFILES;i++) |
| 227 |
nodePercent[i] = nodeAccum[i] / nodeTime; |
| 228 |
|
| 229 |
fprintf(stdout, |
| 230 |
"\n" |
| 231 |
"----------------------------------------------------------------------------\n" |
| 232 |
" Output from Node %d: \n" |
| 233 |
"\n" |
| 234 |
" Time Spent Percent Time Name\n" |
| 235 |
"-------------- ---------------- -----------------------------------------\n", |
| 236 |
j); |
| 237 |
|
| 238 |
for(i=0;i<N_PROFILES;i++){ |
| 239 |
fprintf(stdout, |
| 240 |
" %12G %14G %40s\n", |
| 241 |
nodeAccum[i], |
| 242 |
nodePercent[i], |
| 243 |
theNames[i].myName ); |
| 244 |
} |
| 245 |
|
| 246 |
days = (int)floor( nodeTime / 86400 ); |
| 247 |
donkey = nodeTime - 86400 * days; |
| 248 |
|
| 249 |
hours = (int)floor( donkey / 3600 ); |
| 250 |
donkey -= hours * 3600; |
| 251 |
|
| 252 |
minutes = (int)floor( donkey / 60 ); |
| 253 |
donkey -= minutes * 60; |
| 254 |
|
| 255 |
secs = (int)donkey; |
| 256 |
msecs = (int)( (donkey - secs) * 1000 ); |
| 257 |
|
| 258 |
fprintf( stdout, |
| 259 |
"----------------------------------------------------------------------------\n" |
| 260 |
" Total Time = %03d:%02d:%02d:%02d.%03d ( %G sec )\n", |
| 261 |
days, |
| 262 |
hours, |
| 263 |
minutes, |
| 264 |
secs, |
| 265 |
msecs, |
| 266 |
nodeTime ); |
| 267 |
|
| 268 |
for(i=0;i<N_PROFILES;i++) |
| 269 |
globalAccum[i] += nodeAccum[i]; |
| 270 |
|
| 271 |
globalTime += nodeTime; |
| 272 |
nodeTots[j] = nodeTime; |
| 273 |
} |
| 274 |
|
| 275 |
// print out the totals |
| 276 |
|
| 277 |
for(j=0;j<mpiSim->getNumberProcessors();j++) |
| 278 |
nodePercentTots[j] = nodeTots[j] / globalTime; |
| 279 |
|
| 280 |
for(i=0;i<N_PROFILES;i++) |
| 281 |
globalPercent[i] = globalAccum[i] / globalTime; |
| 282 |
|
| 283 |
fprintf(stdout, |
| 284 |
"\n" |
| 285 |
"----------------------------------------------------------------------------\n" |
| 286 |
" Total Across Nodes\n" |
| 287 |
"\n" |
| 288 |
" Time Spent Percent Time Name\n" |
| 289 |
"-------------- ---------------- -----------------------------------------\n", |
| 290 |
j); |
| 291 |
|
| 292 |
for(i=0;i<N_PROFILES;i++){ |
| 293 |
fprintf(stdout, |
| 294 |
" %12G %14G %40s\n", |
| 295 |
globalAccum[i], |
| 296 |
globalPercent[i], |
| 297 |
theNames[i].myName ); |
| 298 |
} |
| 299 |
fprintf(stdout, |
| 300 |
"\n" |
| 301 |
"\n" ); |
| 302 |
|
| 303 |
for(j=0;j<mpiSim->getNumberProcessors();j++){ |
| 304 |
|
| 305 |
fprintf(stdout, |
| 306 |
" %12G %14G node %d\n", |
| 307 |
nodeTots[j], |
| 308 |
nodePercentTots[j], |
| 309 |
j ); |
| 310 |
} |
| 311 |
|
| 312 |
days = (int)floor( globalTime / 86400 ); |
| 313 |
donkey = nodeTime - 86400 * days; |
| 314 |
|
| 315 |
|
| 316 |
hours = (int)floor( donkey / 3600 ); |
| 317 |
donkey -= hours * 3600; |
| 318 |
|
| 319 |
minutes = (int)floor( donkey / 60 ); |
| 320 |
donkey -= minutes * 60; |
| 321 |
|
| 322 |
secs = (int)donkey; |
| 323 |
msecs = (int)( (donkey - secs) * 1000 ); |
| 324 |
|
| 325 |
fprintf( stdout, |
| 326 |
"----------------------------------------------------------------------------\n" |
| 327 |
" Total Time = %03d:%02d:%02d:%02d.%03d ( %G sec )\n", |
| 328 |
days, |
| 329 |
hours, |
| 330 |
minutes, |
| 331 |
secs, |
| 332 |
msecs, |
| 333 |
globalTime ); |
| 334 |
} |
| 335 |
|
| 336 |
else{ |
| 337 |
|
| 338 |
for(j=1;j<mpiSim->getNumberProcessors();j++){ |
| 339 |
|
| 340 |
if( worldRank == j ){ |
| 341 |
|
| 342 |
MPI_Send( accumTime, N_PROFILES, MPI_DOUBLE, 0, 1, MPI_COMM_WORLD ); |
| 343 |
} |
| 344 |
} |
| 345 |
} |
| 346 |
|
| 347 |
#endif // is_mpi |
| 348 |
|
| 349 |
|
| 350 |
} |
