1 | #include <cassert>
|
---|
2 | #include <cstdlib>
|
---|
3 | #include <cstdio>
|
---|
4 | #include <cstring>
|
---|
5 | #include <climits>
|
---|
6 |
|
---|
7 | extern "C" {
|
---|
8 | #include <locale.h>
|
---|
9 | #include <getopt.h>
|
---|
10 | }
|
---|
11 |
|
---|
12 | #include <unistd.h>
|
---|
13 |
|
---|
14 | #include <chrono>
|
---|
15 |
|
---|
16 | using Clock = std::chrono::high_resolution_clock;
|
---|
17 | using duration_t = std::chrono::duration<double>;
|
---|
18 | using std::chrono::nanoseconds;
|
---|
19 |
|
---|
20 |
|
---|
21 | template<typename Ratio, typename T>
|
---|
22 | T duration_cast(T seconds) {
|
---|
23 | return std::chrono::duration_cast<std::chrono::duration<T, Ratio>>(std::chrono::duration<T>(seconds)).count();
|
---|
24 | }
|
---|
25 |
|
---|
26 | volatile bool run = false;
|
---|
27 | volatile unsigned long long global_counter;
|
---|
28 |
|
---|
29 | #include "libfibre/fibre.h"
|
---|
30 |
|
---|
31 | FibreBarrier * barrier;
|
---|
32 | struct __attribute__((aligned(128))) counter_t {
|
---|
33 | int value = 0;
|
---|
34 | };
|
---|
35 |
|
---|
36 | void fibre_main( counter_t * counter ) {
|
---|
37 | barrier->wait();
|
---|
38 | // /* paranoid */ assert( true == __atomic_load_n(&run, __ATOMIC_RELAXED) );
|
---|
39 |
|
---|
40 | while(__atomic_load_n(&run, __ATOMIC_RELAXED)) {
|
---|
41 | Fibre::forceYield();
|
---|
42 | // fibre_yield();
|
---|
43 | counter->value++;
|
---|
44 | }
|
---|
45 | __atomic_fetch_add(&global_counter, counter->value, __ATOMIC_SEQ_CST);
|
---|
46 | }
|
---|
47 |
|
---|
48 | int main(int argc, char * argv[]) {
|
---|
49 | double duration = 5;
|
---|
50 | int nprocs = 1;
|
---|
51 | int nthreads = 1;
|
---|
52 |
|
---|
53 | std::cout.imbue(std::locale(""));
|
---|
54 | setlocale(LC_ALL, "");
|
---|
55 |
|
---|
56 | for(;;) {
|
---|
57 | static struct option options[] = {
|
---|
58 | {"duration", required_argument, 0, 'd'},
|
---|
59 | {"nprocs", required_argument, 0, 'p'},
|
---|
60 | {"nthreads", required_argument, 0, 't'},
|
---|
61 | {0, 0, 0, 0}
|
---|
62 | };
|
---|
63 |
|
---|
64 | int idx = 0;
|
---|
65 | int opt = getopt_long(argc, argv, "d:p:t:", options, &idx);
|
---|
66 |
|
---|
67 | const char * arg = optarg ? optarg : "";
|
---|
68 | size_t len = 0;
|
---|
69 | char * end;
|
---|
70 | switch(opt) {
|
---|
71 | case -1:
|
---|
72 | goto run;
|
---|
73 | // Numeric Arguments
|
---|
74 | case 'd':
|
---|
75 | duration = strtod(arg, &end);
|
---|
76 | if(*end != '\0') {
|
---|
77 | fprintf(stderr, "Duration must be a valid double, was %s\n", arg);
|
---|
78 | goto usage;
|
---|
79 | }
|
---|
80 | break;
|
---|
81 | case 't':
|
---|
82 | nthreads = strtoul(arg, &end, 10);
|
---|
83 | if(*end != '\0' || nthreads < 1) {
|
---|
84 | fprintf(stderr, "Number of threads must be a positive integer, was %s\n", arg);
|
---|
85 | goto usage;
|
---|
86 | }
|
---|
87 | break;
|
---|
88 | case 'p':
|
---|
89 | nprocs = strtoul(arg, &end, 10);
|
---|
90 | if(*end != '\0' || nprocs < 1) {
|
---|
91 | fprintf(stderr, "Number of processors must be a positive integer, was %s\n", arg);
|
---|
92 | goto usage;
|
---|
93 | }
|
---|
94 | break;
|
---|
95 | // Other cases
|
---|
96 | default: /* ? */
|
---|
97 | fprintf( stderr, "Unkown option '%c'\n", opt);
|
---|
98 | usage:
|
---|
99 | fprintf( stderr, "Usage: %s [options]\n", argv[0]);
|
---|
100 | fprintf( stderr, "\n" );
|
---|
101 | fprintf( stderr, " -d, --duration=DURATION Duration of the experiment, in seconds\n" );
|
---|
102 | fprintf( stderr, " -t, --nthreads=NTHREADS Number of kernel threads\n" );
|
---|
103 | fprintf( stderr, " -q, --nqueues=NQUEUES Number of queues per threads\n" );
|
---|
104 | exit(1);
|
---|
105 | }
|
---|
106 | }
|
---|
107 | run:
|
---|
108 |
|
---|
109 | {
|
---|
110 | printf("Running %d threads on %d processors for %lf seconds\n", nthreads, nprocs, duration);
|
---|
111 |
|
---|
112 | FibreInit();
|
---|
113 | barrier = new FibreBarrier(nthreads + 1);
|
---|
114 | {
|
---|
115 | Context::CurrCluster().addWorkers(nprocs);
|
---|
116 | {
|
---|
117 | counter_t counters[nthreads];
|
---|
118 | Fibre threads[nthreads];
|
---|
119 | for(int i = 0; i < nthreads; i++) {
|
---|
120 | threads[i].run(fibre_main, &counters[i]);
|
---|
121 | }
|
---|
122 | printf("Starting\n");
|
---|
123 | bool is_tty = isatty(STDOUT_FILENO);
|
---|
124 | auto before = Clock::now();
|
---|
125 | run = true;
|
---|
126 |
|
---|
127 | barrier->wait();
|
---|
128 | for(;;) {
|
---|
129 | usleep(500'000);
|
---|
130 | auto now = Clock::now();
|
---|
131 | duration_t durr = now - before;
|
---|
132 | if( durr.count() > duration ) {
|
---|
133 | break;
|
---|
134 | }
|
---|
135 | if(is_tty) {
|
---|
136 | std::cout << "\r" << std::setprecision(4) << durr.count();
|
---|
137 | std::cout.flush();
|
---|
138 | }
|
---|
139 | }
|
---|
140 |
|
---|
141 | auto after = Clock::now();
|
---|
142 | duration_t durr = after - before;
|
---|
143 | duration = durr.count();
|
---|
144 | run = false;
|
---|
145 | printf("\nDone\n");
|
---|
146 | for(auto & thread : threads) {
|
---|
147 | thread.join();
|
---|
148 | }
|
---|
149 |
|
---|
150 | // for(const auto & counter : counters) {
|
---|
151 | // std::cout << counter.value << std::endl;
|
---|
152 | // }
|
---|
153 | }
|
---|
154 | }
|
---|
155 |
|
---|
156 | auto dur_nano = duration_cast<std::nano>(duration);
|
---|
157 |
|
---|
158 | std::cout << "Took " << duration << " s\n";
|
---|
159 | printf("Total yields : %'15llu\n", global_counter );
|
---|
160 | printf("Yields per procs : %'15llu\n", global_counter / nprocs );
|
---|
161 | printf("Yields per second : %'18.2lf\n", ((double)global_counter) / duration );
|
---|
162 | printf("Yields/sec/procs : %'18.2lf\n", (((double)global_counter) / nprocs) / duration );
|
---|
163 | printf("ns per yields : %'18.2lf\n", dur_nano / global_counter );
|
---|
164 | printf("ns per yields/procs : %'18.2lf\n", dur_nano / (global_counter / nprocs) );
|
---|
165 |
|
---|
166 | }
|
---|
167 | }
|
---|