| 1 | #include <stdlib.h>
|
|---|
| 2 | #include <stdio.h>
|
|---|
| 3 | #include <string.h>
|
|---|
| 4 | #include <limits.h>
|
|---|
| 5 |
|
|---|
| 6 | extern "C" {
|
|---|
| 7 | #include <locale.h>
|
|---|
| 8 | #include <getopt.h>
|
|---|
| 9 | }
|
|---|
| 10 |
|
|---|
| 11 | #include <unistd.h>
|
|---|
| 12 |
|
|---|
| 13 | #include <clock.hfa>
|
|---|
| 14 | #include <time.hfa>
|
|---|
| 15 |
|
|---|
| 16 | #include "../benchcltr.hfa"
|
|---|
| 17 |
|
|---|
| 18 | extern bool traceHeapOn();
|
|---|
| 19 |
|
|---|
| 20 |
|
|---|
| 21 | volatile bool run = false;
|
|---|
| 22 | volatile unsigned long long global_counter;
|
|---|
| 23 |
|
|---|
| 24 | thread __attribute__((aligned(64))) Yielder {
|
|---|
| 25 | unsigned long long counter;
|
|---|
| 26 | };
|
|---|
| 27 | void ?{}( Yielder & this ) {
|
|---|
| 28 | this.counter = 0;
|
|---|
| 29 | ((thread&)this){ "Yielder Thread", *the_benchmark_cluster };
|
|---|
| 30 | }
|
|---|
| 31 |
|
|---|
| 32 | void main( Yielder & this ) {
|
|---|
| 33 | park( __cfaabi_dbg_ctx );
|
|---|
| 34 | /* paranoid */ assert( true == __atomic_load_n(&run, __ATOMIC_RELAXED) );
|
|---|
| 35 |
|
|---|
| 36 | while(__atomic_load_n(&run, __ATOMIC_RELAXED)) {
|
|---|
| 37 | yield();
|
|---|
| 38 | this.counter++;
|
|---|
| 39 | }
|
|---|
| 40 | __atomic_fetch_add(&global_counter, this.counter, __ATOMIC_SEQ_CST);
|
|---|
| 41 | }
|
|---|
| 42 |
|
|---|
| 43 | int main(int argc, char * argv[]) {
|
|---|
| 44 | double duration = 5;
|
|---|
| 45 | int nprocs = 1;
|
|---|
| 46 | int nthreads = 1;
|
|---|
| 47 |
|
|---|
| 48 |
|
|---|
| 49 |
|
|---|
| 50 | for(;;) {
|
|---|
| 51 | static struct option options[] = {
|
|---|
| 52 | {"duration", required_argument, 0, 'd'},
|
|---|
| 53 | {"nprocs", required_argument, 0, 'p'},
|
|---|
| 54 | {"nthreads", required_argument, 0, 't'},
|
|---|
| 55 | {0, 0, 0, 0}
|
|---|
| 56 | };
|
|---|
| 57 |
|
|---|
| 58 | int idx = 0;
|
|---|
| 59 | int opt = getopt_long(argc, argv, "d:p:t:", options, &idx);
|
|---|
| 60 |
|
|---|
| 61 | char * arg = optarg ? optarg : "";
|
|---|
| 62 | size_t len = 0;
|
|---|
| 63 | char * end;
|
|---|
| 64 | switch(opt) {
|
|---|
| 65 | case -1:
|
|---|
| 66 | goto run;
|
|---|
| 67 | // Numeric Arguments
|
|---|
| 68 | case 'd':
|
|---|
| 69 | duration = strtod(arg, &end);
|
|---|
| 70 | if(*end != '\0') {
|
|---|
| 71 | fprintf(stderr, "Duration must be a valid double, was %s\n", arg);
|
|---|
| 72 | goto usage;
|
|---|
| 73 | }
|
|---|
| 74 | break;
|
|---|
| 75 | case 't':
|
|---|
| 76 | nthreads = strtoul(arg, &end, 10);
|
|---|
| 77 | if(*end != '\0' || nthreads < 1) {
|
|---|
| 78 | fprintf(stderr, "Number of threads must be a positive integer, was %s\n", arg);
|
|---|
| 79 | goto usage;
|
|---|
| 80 | }
|
|---|
| 81 | break;
|
|---|
| 82 | case 'p':
|
|---|
| 83 | nprocs = strtoul(arg, &end, 10);
|
|---|
| 84 | if(*end != '\0' || nprocs < 1) {
|
|---|
| 85 | fprintf(stderr, "Number of processors must be a positive integer, was %s\n", arg);
|
|---|
| 86 | goto usage;
|
|---|
| 87 | }
|
|---|
| 88 | break;
|
|---|
| 89 | // Other cases
|
|---|
| 90 | default: /* ? */
|
|---|
| 91 | fprintf( stderr, "Unkown option '%c'\n", opt);
|
|---|
| 92 | usage:
|
|---|
| 93 | fprintf( stderr, "Usage: %s [options]\n", argv[0]);
|
|---|
| 94 | fprintf( stderr, "\n" );
|
|---|
| 95 | fprintf( stderr, " -d, --duration=DURATION Duration of the experiment, in seconds\n" );
|
|---|
| 96 | fprintf( stderr, " -t, --nthreads=NTHREADS Number of kernel threads\n" );
|
|---|
| 97 | fprintf( stderr, " -q, --nqueues=NQUEUES Number of queues per threads\n" );
|
|---|
| 98 | exit(1);
|
|---|
| 99 | }
|
|---|
| 100 | }
|
|---|
| 101 | run:
|
|---|
| 102 |
|
|---|
| 103 | {
|
|---|
| 104 | printf("Running %d threads on %d processors for %lf seconds\n", nthreads, nprocs, duration);
|
|---|
| 105 |
|
|---|
| 106 | Time start, end;
|
|---|
| 107 | BenchCluster cl;
|
|---|
| 108 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 109 | print_stats_at_exit( cl.self );
|
|---|
| 110 | #endif
|
|---|
| 111 | {
|
|---|
| 112 | BenchProc procs[nprocs];
|
|---|
| 113 | {
|
|---|
| 114 | Yielder threads[nthreads];
|
|---|
| 115 | printf("Starting\n");
|
|---|
| 116 | start = getTime();
|
|---|
| 117 | run = true;
|
|---|
| 118 |
|
|---|
| 119 | bool is_tty = isatty(STDOUT_FILENO);
|
|---|
| 120 | for(i; nthreads) {
|
|---|
| 121 | unpark( threads[i] __cfaabi_dbg_ctx2 );
|
|---|
| 122 | }
|
|---|
| 123 | for() {
|
|---|
| 124 | sleep(500`ms);
|
|---|
| 125 | end = getTime();
|
|---|
| 126 | if( (end - start) >= duration`s ) {
|
|---|
| 127 | break;
|
|---|
| 128 | }
|
|---|
| 129 | if(is_tty) {
|
|---|
| 130 | printf("\r%.4lf", (double)(end - start)`s);
|
|---|
| 131 | fflush(stdout);
|
|---|
| 132 | }
|
|---|
| 133 | }
|
|---|
| 134 |
|
|---|
| 135 | run = false;
|
|---|
| 136 | end = getTime();
|
|---|
| 137 | printf("Done\n");
|
|---|
| 138 | }
|
|---|
| 139 | }
|
|---|
| 140 |
|
|---|
| 141 | printf("Took %'ld ms\n", (end - start)`ms);
|
|---|
| 142 | printf("Total yields : %'15llu\n", global_counter);
|
|---|
| 143 | printf("Yields per procs : %'15llu\n", global_counter / nprocs);
|
|---|
| 144 | printf("Yields per second : %'18.2lf\n", ((double)global_counter) / (end - start)`s);
|
|---|
| 145 | printf("Yields/sec/procs : %'18.2lf\n", (((double)global_counter) / nprocs) / (end - start)`s);
|
|---|
| 146 | printf("ns per yields : %'18.2lf\n", ((double)(end - start)`ns) / global_counter);
|
|---|
| 147 | printf("ns per yields/procs : %'18.2lf\n", ((double)(end - start)`ns) / (global_counter / nprocs));
|
|---|
| 148 |
|
|---|
| 149 | }
|
|---|
| 150 | }
|
|---|