| 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 | #include <stats.hfa>
|
|---|
| 16 |
|
|---|
| 17 | #include "../benchcltr.hfa"
|
|---|
| 18 |
|
|---|
| 19 | extern bool traceHeapOn();
|
|---|
| 20 |
|
|---|
| 21 |
|
|---|
| 22 | volatile bool run = false;
|
|---|
| 23 | volatile unsigned long long global_counter;
|
|---|
| 24 |
|
|---|
| 25 | thread __attribute__((aligned(128))) Yielder {
|
|---|
| 26 | unsigned long long counter;
|
|---|
| 27 | };
|
|---|
| 28 | void ?{}( Yielder & this ) {
|
|---|
| 29 | this.counter = 0;
|
|---|
| 30 | ((thread&)this){ "Yielder Thread", *the_benchmark_cluster };
|
|---|
| 31 | }
|
|---|
| 32 |
|
|---|
| 33 | void main( Yielder & this ) {
|
|---|
| 34 | park();
|
|---|
| 35 | /* paranoid */ assert( true == __atomic_load_n(&run, __ATOMIC_RELAXED) );
|
|---|
| 36 |
|
|---|
| 37 | while(__atomic_load_n(&run, __ATOMIC_RELAXED)) {
|
|---|
| 38 | yield();
|
|---|
| 39 | this.counter++;
|
|---|
| 40 | }
|
|---|
| 41 | __atomic_fetch_add(&global_counter, this.counter, __ATOMIC_SEQ_CST);
|
|---|
| 42 | }
|
|---|
| 43 |
|
|---|
| 44 | int main(int argc, char * argv[]) {
|
|---|
| 45 | unsigned num_io = 1;
|
|---|
| 46 | io_context_params params;
|
|---|
| 47 |
|
|---|
| 48 | cfa_option opt[] = {
|
|---|
| 49 | BENCH_OPT_CFA
|
|---|
| 50 | };
|
|---|
| 51 | int opt_cnt = sizeof(opt) / sizeof(cfa_option);
|
|---|
| 52 |
|
|---|
| 53 | char **left;
|
|---|
| 54 | parse_args( argc, argv, opt, opt_cnt, "[OPTIONS]...\ncforall yield benchmark", left );
|
|---|
| 55 |
|
|---|
| 56 | {
|
|---|
| 57 | printf("Running %d threads on %d processors for %f seconds\n", nthreads, nprocs, duration);
|
|---|
| 58 |
|
|---|
| 59 | Time start, end;
|
|---|
| 60 | BenchCluster cl = { num_io, params, CFA_STATS_READY_Q };
|
|---|
| 61 | {
|
|---|
| 62 | BenchProc procs[nprocs];
|
|---|
| 63 | {
|
|---|
| 64 | Yielder threads[nthreads];
|
|---|
| 65 | printf("Starting\n");
|
|---|
| 66 |
|
|---|
| 67 | bool is_tty = isatty(STDOUT_FILENO);
|
|---|
| 68 | start = timeHiRes();
|
|---|
| 69 | run = true;
|
|---|
| 70 |
|
|---|
| 71 | for(i; nthreads) {
|
|---|
| 72 | unpark( threads[i] );
|
|---|
| 73 | }
|
|---|
| 74 | wait(duration, start, end, is_tty);
|
|---|
| 75 |
|
|---|
| 76 | run = false;
|
|---|
| 77 | end = timeHiRes();
|
|---|
| 78 | printf("\nDone\n");
|
|---|
| 79 | }
|
|---|
| 80 | }
|
|---|
| 81 |
|
|---|
| 82 | printf("Took %'ld ms\n", (end - start)`ms);
|
|---|
| 83 | printf("Yields per second : %'18.2lf\n", ((double)global_counter) / (end - start)`s);
|
|---|
| 84 | printf("ns per yields : %'18.2lf\n", ((double)(end - start)`ns) / global_counter);
|
|---|
| 85 | printf("Total yields : %'15llu\n", global_counter);
|
|---|
| 86 | printf("Yields per procs : %'15llu\n", global_counter / nprocs);
|
|---|
| 87 | printf("Yields/sec/procs : %'18.2lf\n", (((double)global_counter) / nprocs) / (end - start)`s);
|
|---|
| 88 | printf("ns per yields/procs : %'18.2lf\n", ((double)(end - start)`ns) / (global_counter / nprocs));
|
|---|
| 89 | fflush(stdout);
|
|---|
| 90 | }
|
|---|
| 91 | }
|
|---|