1 2 #include <sys/time.h> 3 /* int gettimeofday(struct timeval *tp, struct timezone *tzp); */ 4 5 double second() 6 { 7 /* struct timeval { long tv_sec; 8 long tv_usec; }; 9 10 struct timezone { int tz_minuteswest; 11 int tz_dsttime; }; */ 12 13 struct timeval tp; 14 struct timezone tzp; 15 16 (void)gettimeofday(&tp,&tzp); 17 return ((double) tp.tv_sec + (double) tp.tv_usec * 1.e-6); 18 } 19 # include <stdio.h> 20 # include <math.h> 21 # include <limits.h> 22 # include <float.h> 23 # include <sys/time.h> 24 25 /* 26 * Program: Stream 27 * Programmer: Joe R. Zagar 28 * Revision: 4.0-BETA, October 24, 1995 29 * Original code developed by John D. McCalpin 30 * 31 * This program measures memory transfer rates in MB/s for simple 32 * computational kernels coded in C. These numbers reveal the quality 33 * of code generation for simple uncacheable kernels as well as showing 34 * the cost of floating-point operations relative to memory accesses. 35 * 36 * INSTRUCTIONS: 37 * 38 * 1) Stream requires a good bit of memory to run. Adjust the 39 * value of 'N' (below) to give a 'timing calibration' of 40 * at least 20 clock-ticks. This will provide rate estimates 41 * that should be good to about 5% precision. 42 */ 43 44 # define N 2000000 45 # define NTIMES 50 46 # define OFFSET 0 47 48 /* 49 * 3) Compile the code with full optimization. Many compilers 50 * generate unreasonably bad code before the optimizer tightens 51 * things up. If the results are unreasonably good, on the 52 * other hand, the optimizer might be too smart for me! 53 * 54 * Try compiling with: 55 * cc -O stream_d.c second.c -o stream_d -lm 56 * 57 * This is known to work on Cray, SGI, IBM, and Sun machines. 58 * 59 * 60 * 4) Mail the results to mccalpin@cs.virginia.edu 61 * Be sure to include: 62 * a) computer hardware model number and software revision 63 * b) the compiler flags 64 * c) all of the output from the test case. 65 * Thanks! 66 * 67 */ 68 69 # define HLINE "-------------------------------------------------------------\n" 70 71 # ifndef MIN 72 # define MIN(x,y) ((x)<(y) ? (x) : (y)) 73 # endif 74 # ifndef MAX 75 # define MAX(x,y) ((x)>(y) ? (x) : (y)) 76 # endif 77 78 static double a[N+OFFSET], 79 b[N+OFFSET], 80 c[N+OFFSET]; 81 /*double *a,*b,*c;*/ 82 83 static double mintime[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX}; 84 85 static const char *label[4] = {"Copy: ", "Scale: ", "Add: ", "Triad: "}; 86 87 static double bytes[4] = { 88 2 * sizeof(double) * N, 89 2 * sizeof(double) * N, 90 3 * sizeof(double) * N, 91 3 * sizeof(double) * N 92 }; 93 94 extern double second(); 95 96 #include <mpi.h> 97 98 int main(int argc,char **args) 99 { 100 int quantum, checktick(); 101 register int j, k; 102 double scalar, t, times[4][NTIMES],irate[4],rate[4]; 103 int rank,size,resultlen; 104 char hostname[MPI_MAX_PROCESSOR_NAME]; 105 106 MPI_Init(&argc,&args); 107 MPI_Comm_rank(MPI_COMM_WORLD,&rank); 108 MPI_Comm_size(MPI_COMM_WORLD,&size); 109 if (!rank) printf("Number of MPI processes %d\n",size); 110 111 for (j=0; j<size; j++) { 112 if (rank == j) { 113 MPI_Get_processor_name(hostname,&resultlen); 114 printf("Process %d %s\n",rank,hostname); 115 fflush(stdout); 116 } 117 MPI_Barrier(MPI_COMM_WORLD); 118 } 119 120 /* --- SETUP --- determine precision and check timing --- */ 121 122 if (!rank) { 123 /*printf(HLINE); 124 printf("Array size = %d, Offset = %d\n" , N, OFFSET); 125 printf("Total memory required = %.1f MB.\n", (3 * N * BytesPerWord) / 1048576.0); 126 printf("Each test is run %d times, but only\n", NTIMES); 127 printf("the *best* time for each is used.\n"); 128 printf(HLINE); */ 129 } 130 131 /* Get initial value for system clock. */ 132 133 /* a = malloc(N*sizeof(double)); 134 b = malloc(N*sizeof(double)); 135 c = malloc(N*sizeof(double));*/ 136 for (j=0; j<N; j++) { 137 a[j] = 1.0; 138 b[j] = 2.0; 139 c[j] = 0.0; 140 } 141 142 if (!rank) { 143 if ((quantum = checktick()) >= 1) ; /* printf("Your clock granularity/precision appears to be %d microseconds.\n", quantum); */ 144 else ; /* printf("Your clock granularity appears to be less than one microsecond.\n");*/ 145 } 146 147 t = second(); 148 for (j = 0; j < N; j++) a[j] = 2.0E0 * a[j]; 149 t = 1.0E6 * (second() - t); 150 151 if (!rank) { 152 /* printf("Each test below will take on the order of %d microseconds.\n", (int) t); 153 printf(" (= %d clock ticks)\n", (int) (t/quantum)); 154 printf("Increase the size of the arrays if this shows that\n"); 155 printf("you are not getting at least 20 clock ticks per test.\n"); 156 printf(HLINE);*/ 157 } 158 159 160 /* --- MAIN LOOP --- repeat test cases NTIMES times --- */ 161 162 scalar = 3.0; 163 for (k=0; k<NTIMES; k++) 164 { 165 MPI_Barrier(MPI_COMM_WORLD); 166 times[0][k] = second(); 167 /* should all these barriers be pulled outside of the time call? */ 168 MPI_Barrier(MPI_COMM_WORLD); 169 for (j=0; j<N; j++) c[j] = a[j]; 170 MPI_Barrier(MPI_COMM_WORLD); 171 times[0][k] = second() - times[0][k]; 172 173 times[1][k] = second(); 174 MPI_Barrier(MPI_COMM_WORLD); 175 for (j=0; j<N; j++) b[j] = scalar*c[j]; 176 MPI_Barrier(MPI_COMM_WORLD); 177 times[1][k] = second() - times[1][k]; 178 179 times[2][k] = second(); 180 MPI_Barrier(MPI_COMM_WORLD); 181 for (j=0; j<N; j++) c[j] = a[j]+b[j]; 182 MPI_Barrier(MPI_COMM_WORLD); 183 times[2][k] = second() - times[2][k]; 184 185 times[3][k] = second(); 186 MPI_Barrier(MPI_COMM_WORLD); 187 for (j=0; j<N; j++) a[j] = b[j]+scalar*c[j]; 188 MPI_Barrier(MPI_COMM_WORLD); 189 times[3][k] = second() - times[3][k]; 190 } 191 192 /* --- SUMMARY --- */ 193 194 for (k=0; k<NTIMES; k++) 195 for (j=0; j<4; j++) mintime[j] = MIN(mintime[j], times[j][k]); 196 197 for (j=0; j<4; j++) irate[j] = 1.0E-06 * bytes[j]/mintime[j]; 198 MPI_Reduce(irate,rate,4,MPI_DOUBLE,MPI_SUM,0,MPI_COMM_WORLD); 199 200 if (!rank) { 201 printf("Function Rate (MB/s) \n"); 202 for (j=0; j<4; j++) printf("%s%11.4f\n", label[j],rate[j]); 203 } 204 MPI_Finalize(); 205 return 0; 206 } 207 208 # define M 20 209 210 int checktick() 211 { 212 int i, minDelta, Delta; 213 double t1, t2, timesfound[M]; 214 215 /* Collect a sequence of M unique time values from the system. */ 216 217 for (i = 0; i < M; i++) { 218 t1 = second(); 219 while (((t2=second()) - t1) < 1.0E-6) ; 220 timesfound[i] = t1 = t2; 221 } 222 223 /* 224 * Determine the minimum difference between these M values. 225 * This result will be our estimate (in microseconds) for the 226 * clock granularity. 227 */ 228 229 minDelta = 1000000; 230 for (i = 1; i < M; i++) { 231 Delta = (int)(1.0E6 * (timesfound[i]-timesfound[i-1])); 232 minDelta = MIN(minDelta, MAX(Delta,0)); 233 } 234 235 return(minDelta); 236 } 237 238