xref: /petsc/src/benchmarks/streams/MPIVersion.c (revision 0dad66eaf1c61c5c8276872bd6fceedd70f85537)
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   MPI_Status   status;
106 
107   MPI_Init(&argc,&args);
108   MPI_Comm_rank(MPI_COMM_WORLD,&rank);
109   MPI_Comm_size(MPI_COMM_WORLD,&size);
110   if (!rank) printf("Number of MPI processes %d\n",size);
111 
112   MPI_Get_processor_name(hostname,&resultlen);
113   if (!rank) {
114     printf("Process %d %s\n",rank,hostname);
115     for (j=1; j<size; j++) {
116       MPI_Recv(hostname,MPI_MAX_PROCESSOR_NAME,MPI_CHAR,j,0,MPI_COMM_WORLD,&status);
117       printf("Process %d %s\n",j,hostname);
118     }
119  } else {
120    MPI_Send(hostname,MPI_MAX_PROCESSOR_NAME,MPI_CHAR,0,0,MPI_COMM_WORLD);
121  }
122  MPI_Barrier(MPI_COMM_WORLD);
123 
124   /* --- SETUP --- determine precision and check timing --- */
125 
126   if (!rank) {
127     /*printf(HLINE);
128     printf("Array size = %d, Offset = %d\n" , N, OFFSET);
129     printf("Total memory required = %.1f MB.\n", (3 * N * BytesPerWord) / 1048576.0);
130     printf("Each test is run %d times, but only\n", NTIMES);
131     printf("the *best* time for each is used.\n");
132     printf(HLINE); */
133   }
134 
135   /* Get initial value for system clock. */
136 
137   /*  a = malloc(N*sizeof(double));
138   b = malloc(N*sizeof(double));
139   c = malloc(N*sizeof(double));*/
140   for (j=0; j<N; j++) {
141     a[j] = 1.0;
142     b[j] = 2.0;
143     c[j] = 0.0;
144   }
145 
146   if (!rank) {
147     if  ((quantum = checktick()) >= 1) ; /* printf("Your clock granularity/precision appears to be %d microseconds.\n", quantum); */
148     else ; /* printf("Your clock granularity appears to be less than one microsecond.\n");*/
149   }
150 
151   t = second();
152   for (j = 0; j < N; j++) a[j] = 2.0E0 * a[j];
153   t = 1.0E6 * (second() - t);
154 
155   if (!rank) {
156     /*  printf("Each test below will take on the order of %d microseconds.\n", (int) t);
157     printf("   (= %d clock ticks)\n", (int) (t/quantum));
158     printf("Increase the size of the arrays if this shows that\n");
159     printf("you are not getting at least 20 clock ticks per test.\n");
160     printf(HLINE);*/
161   }
162 
163 
164   /*   --- MAIN LOOP --- repeat test cases NTIMES times --- */
165 
166   scalar = 3.0;
167   for (k=0; k<NTIMES; k++)
168   {
169     MPI_Barrier(MPI_COMM_WORLD);
170     times[0][k] = second();
171     /* should all these barriers be pulled outside of the time call? */
172     MPI_Barrier(MPI_COMM_WORLD);
173     for (j=0; j<N; j++) c[j] = a[j];
174     MPI_Barrier(MPI_COMM_WORLD);
175     times[0][k] = second() - times[0][k];
176 
177     times[1][k] = second();
178     MPI_Barrier(MPI_COMM_WORLD);
179     for (j=0; j<N; j++) b[j] = scalar*c[j];
180     MPI_Barrier(MPI_COMM_WORLD);
181     times[1][k] = second() - times[1][k];
182 
183     times[2][k] = second();
184     MPI_Barrier(MPI_COMM_WORLD);
185     for (j=0; j<N; j++) c[j] = a[j]+b[j];
186     MPI_Barrier(MPI_COMM_WORLD);
187     times[2][k] = second() - times[2][k];
188 
189     times[3][k] = second();
190     MPI_Barrier(MPI_COMM_WORLD);
191     for (j=0; j<N; j++) a[j] = b[j]+scalar*c[j];
192     MPI_Barrier(MPI_COMM_WORLD);
193     times[3][k] = second() - times[3][k];
194   }
195 
196   /*   --- SUMMARY --- */
197 
198   for (k=0; k<NTIMES; k++)
199     for (j=0; j<4; j++) mintime[j] = MIN(mintime[j], times[j][k]);
200 
201   for (j=0; j<4; j++) irate[j] = 1.0E-06 * bytes[j]/mintime[j];
202   MPI_Reduce(irate,rate,4,MPI_DOUBLE,MPI_SUM,0,MPI_COMM_WORLD);
203 
204   if (!rank) {
205     printf("Function      Rate (MB/s) \n");
206     for (j=0; j<4; j++) printf("%s%11.4f\n", label[j],rate[j]);
207   }
208   MPI_Finalize();
209   return 0;
210 }
211 
212 # define        M        20
213 
214 int checktick()
215 {
216   int    i, minDelta, Delta;
217   double t1, t2, timesfound[M];
218 
219 /*  Collect a sequence of M unique time values from the system. */
220 
221   for (i = 0; i < M; i++) {
222     t1 = second();
223     while (((t2=second()) - t1) < 1.0E-6) ;
224     timesfound[i] = t1 = t2;
225   }
226 
227 /*
228 * Determine the minimum difference between these M values.
229 * This result will be our estimate (in microseconds) for the
230 * clock granularity.
231 */
232 
233   minDelta = 1000000;
234   for (i = 1; i < M; i++) {
235     Delta    = (int)(1.0E6 * (timesfound[i]-timesfound[i-1]));
236     minDelta = MIN(minDelta, MAX(Delta,0));
237   }
238 
239   return(minDelta);
240 }
241 
242