test/sysperf/randread.c

   1 /*
   2  * cc randread.c -o ~/bin/randread -O2 -lm
   3  *
   4  * randread device [bufsize:512 [range%:90 [nprocs:32]]]
   5  *
   6  * requires TSC
   7  */
   8 #include <sys/types.h>
   9 #include <sys/sysctl.h>
  10 #include <sys/stat.h>
  11 #include <sys/file.h>
  12 #include <sys/mman.h>
  13 #include <sys/errno.h>
  14 #include <sys/wait.h>
  15 #include <string.h>
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <unistd.h>
  19 #include <math.h>
  20 #include <assert.h>
  21 #include <machine/atomic.h>
  22 #include <machine/cpufunc.h>
  23
  24 typedef struct pdata {
  25         int64_t counter;
  26         int64_t lotime;
  27         int64_t hitime;
  28         int64_t tsc_total1;
  29         int64_t tsc_total2;
  30         int64_t unused00;
  31         int64_t unused01;
  32         int     unused02;
  33         int     reset;
  34 } pdata_t;
  35
  36 int
  37 main(int ac, char **av)
  38 {
  39     char *buf;
  40     size_t bytes = 512;
  41     off_t limit;
  42     int fd;
  43     int i;
  44     int loops;
  45     int nprocs = 32;
  46     double range = 90.0;
  47     volatile pdata_t *pdata;
  48     int64_t tsc1;
  49     int64_t tsc2;
  50     int64_t delta;
  51     int64_t tscfreq = 0;
  52     int64_t lotime;
  53     int64_t hitime;
  54     size_t tscfreq_size = sizeof(tscfreq);
  55
  56     sysctlbyname("hw.tsc_frequency", &tscfreq, &tscfreq_size, NULL, 0);
  57     assert(tscfreq != 0);
  58
  59     if (ac < 2 || ac > 5) {
  60         fprintf(stderr, "%s <device> [bufsize:512 [range%:90 [nprocs:32]]]\n",
  61                 av[0]);
  62         exit (1);
  63     }
  64
  65     if (ac >= 3) {
  66         bytes = (size_t)strtoul(av[2], NULL, 0);
  67         if (bytes < 512 || (bytes ^ (bytes - 1)) != ((bytes << 1) - 1)) {
  68             fprintf(stderr, "bytes must be a power of 2 >= 512\n");
  69             exit (1);
  70         }
  71     }
  72     buf = malloc(bytes);
  73
  74     if (ac >= 4) {
  75         range = strtod(av[3], NULL);
  76     }
  77
  78     if (ac >= 5) {
  79         nprocs = strtol(av[4], NULL, 0);
  80         if (nprocs < 0 || nprocs > 512) {
  81             fprintf(stderr, "absurd nprocs (%d)\n", nprocs);
  82             exit(1);
  83         }
  84     }
  85
  86     fd = open(av[1], O_RDONLY);
  87     if (fd < 0) {
  88         fprintf(stderr, "open %s: %s\n", av[1], strerror(errno));
  89         exit (1);
  90     }
  91
  92     lseek(fd, 0L, 2);
  93     limit = lseek(fd, 0L, 1);
  94     limit = (off_t)((double)limit * range / 100.0);
  95     limit &= ~(off_t)(bytes - 1);
  96     printf("device %s bufsize %zd limit %4.3fGB nprocs %d\n",
  97         av[1], bytes, (double)limit / (1024.0*1024.0*1024.0), nprocs);
  98
  99     pdata = mmap(NULL, nprocs * sizeof(*pdata), PROT_READ|PROT_WRITE,
 100                     MAP_SHARED|MAP_ANON, -1, 0);
 101
 102     for (i = 0; i < nprocs; ++i) {
 103         if (fork() == 0) {
 104             close(fd);
 105             fd = open(av[1], O_RDONLY);
 106             srandomdev();
 107             pdata += i;
 108
 109             tsc2 = rdtsc();
 110             pdata->lotime = 0x7FFFFFFFFFFFFFFFLL;
 111
 112             for (;;) {
 113                 long pos;
 114
 115                 if (pdata->reset) {
 116                         pdata->counter = 0;
 117                         pdata->tsc_total1 = 0;
 118                         pdata->tsc_total2 = 0;
 119                         pdata->lotime = 0x7FFFFFFFFFFFFFFFLL;
 120                         pdata->hitime = 0;
 121                         pdata->reset = 0;
 122                 }
 123
 124                 pos = random() ^ ((long)random() << 31);
 125                 pos &= 0x7FFFFFFFFFFFFFFFLLU;
 126                 pos = (pos % limit) & ~(off_t)(bytes - 1);
 127                 lseek(fd, pos, 0);
 128                 read(fd, buf, bytes);
 129                 tsc1 = tsc2;
 130                 tsc2 = rdtsc();
 131                 delta = tsc2 - tsc1;
 132                 ++pdata->counter;
 133                 pdata->tsc_total1 += delta;
 134                 pdata->tsc_total2 += delta * delta;
 135                 if (pdata->lotime > delta)
 136                         pdata->lotime = delta;
 137                 if (pdata->hitime < delta)
 138                         pdata->hitime = delta;
 139             }
 140         }
 141     }
 142
 143     tsc2 = rdtsc();
 144     loops = 0;
 145
 146     for (;;) {
 147         int64_t count;
 148         int64_t total1;
 149         int64_t total2;
 150         double v;
 151         double lo;
 152         double hi;
 153         double s1;
 154         double s2;
 155         double stddev;
 156
 157         sleep(1);
 158         lotime = pdata[0].lotime;
 159         hitime = pdata[0].hitime;
 160         total1 = 0;
 161         total2 = 0;
 162         count = 0;
 163
 164         for (i = 0; i < nprocs; ++i) {
 165                 count += pdata[i].counter;
 166                 total1 += pdata[i].tsc_total1;
 167                 total2 += pdata[i].tsc_total2;
 168                 if (lotime > pdata[i].lotime)
 169                         lotime = pdata[i].lotime;
 170                 if (hitime < pdata[i].hitime)
 171                         hitime = pdata[i].hitime;
 172                 pdata[i].reset = 1;
 173         }
 174         tsc1 = tsc2;
 175         tsc2 = rdtsc();
 176         delta = tsc2 - tsc1;
 177         v = count * ((double)delta / (double)tscfreq);
 178         lo = (double)lotime / (double)tscfreq;
 179         hi = (double)hitime / (double)tscfreq;
 180
 181         s1 = ((double)total2 - (double)total1 * (double)total1 / (double)count) / ((double)count - 1);
 182         if (s1 < 0.0)
 183                 stddev = -sqrt(-s1);
 184         else
 185                 stddev = sqrt(s1);
 186         stddev = stddev / (double)tscfreq;      /* normalize to 1 second units */
 187
 188         if (loops) {
 189                 printf("%6.0f/s avg=%6.2fuS bw=%-6.2fMB/s "
 190                        "lo=%-3.2fuS, hi=%-3.2fuS stddev=%3.2fuS\n",
 191                        v,
 192                        1e6 * nprocs / v,
 193                        (double)count * bytes / 1e6 / ((double)delta / (double)tscfreq),
 194                        lo * 1e6,
 195                        hi * 1e6,
 196                        stddev * 1e6);
 197         }
 198         ++loops;
 199     }
 200     return 0;
 201 }