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( __cfaabi_dbg_ctx ); |
---|
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 | double duration = 5; |
---|
46 | int nprocs = 1; |
---|
47 | int nthreads = 1; |
---|
48 | bool silent = false; |
---|
49 | bool procstats = false; |
---|
50 | |
---|
51 | for(;;) { |
---|
52 | static struct option options[] = { |
---|
53 | {"duration", required_argument, 0, 'd'}, |
---|
54 | {"nprocs", required_argument, 0, 'p'}, |
---|
55 | {"nthreads", required_argument, 0, 't'}, |
---|
56 | {"nostats", no_argument , 0, 'S'}, |
---|
57 | {"procstat", no_argument , 0, 'P'}, |
---|
58 | {0, 0, 0, 0} |
---|
59 | }; |
---|
60 | |
---|
61 | int idx = 0; |
---|
62 | int opt = getopt_long(argc, argv, "d:p:t:SP", options, &idx); |
---|
63 | |
---|
64 | char * arg = optarg ? optarg : ""; |
---|
65 | size_t len = 0; |
---|
66 | char * end; |
---|
67 | switch(opt) { |
---|
68 | case -1: |
---|
69 | goto run; |
---|
70 | // Numeric Arguments |
---|
71 | case 'd': |
---|
72 | duration = strtod(arg, &end); |
---|
73 | if(*end != '\0') { |
---|
74 | fprintf(stderr, "Duration must be a valid double, was %s\n", arg); |
---|
75 | goto usage; |
---|
76 | } |
---|
77 | break; |
---|
78 | case 't': |
---|
79 | nthreads = strtoul(arg, &end, 10); |
---|
80 | if(*end != '\0' || nthreads < 1) { |
---|
81 | fprintf(stderr, "Number of threads must be a positive integer, was %s\n", arg); |
---|
82 | goto usage; |
---|
83 | } |
---|
84 | break; |
---|
85 | case 'p': |
---|
86 | nprocs = strtoul(arg, &end, 10); |
---|
87 | if(*end != '\0' || nprocs < 1) { |
---|
88 | fprintf(stderr, "Number of processors must be a positive integer, was %s\n", arg); |
---|
89 | goto usage; |
---|
90 | } |
---|
91 | break; |
---|
92 | case 'S': |
---|
93 | silent = true; |
---|
94 | break; |
---|
95 | case 'P': |
---|
96 | procstats = true; |
---|
97 | break; |
---|
98 | // Other cases |
---|
99 | default: /* ? */ |
---|
100 | fprintf( stderr, "Unkown option '%c'\n", opt); |
---|
101 | usage: |
---|
102 | fprintf( stderr, "Usage: %s [options]\n", argv[0]); |
---|
103 | fprintf( stderr, "\n" ); |
---|
104 | fprintf( stderr, " -d, --duration=DURATION Duration of the experiment, in seconds\n" ); |
---|
105 | fprintf( stderr, " -t, --nthreads=NTHREADS Number of kernel threads\n" ); |
---|
106 | fprintf( stderr, " -q, --nqueues=NQUEUES Number of queues per threads\n" ); |
---|
107 | fprintf( stderr, " -S, --nostats Don't print cluster stats\n" ); |
---|
108 | fprintf( stderr, " -P, --procstat Print processor stats" ); |
---|
109 | exit(1); |
---|
110 | } |
---|
111 | } |
---|
112 | run: |
---|
113 | |
---|
114 | { |
---|
115 | printf("Running %d threads on %d processors for %lf seconds\n", nthreads, nprocs, duration); |
---|
116 | |
---|
117 | Time start, end; |
---|
118 | BenchCluster cl = { 0 }; |
---|
119 | #if !defined(__CFA_NO_STATISTICS__) |
---|
120 | if( !silent ) { |
---|
121 | print_stats_at_exit( cl.self, CFA_STATS_READY_Q ); |
---|
122 | } |
---|
123 | #endif |
---|
124 | { |
---|
125 | BenchProc procs[nprocs]; |
---|
126 | #if !defined(__CFA_NO_STATISTICS__) |
---|
127 | if( procstats ) { |
---|
128 | for(i; nprocs) { |
---|
129 | print_stats_at_exit( procs[i].self, CFA_STATS_READY_Q ); |
---|
130 | } |
---|
131 | } |
---|
132 | #endif |
---|
133 | { |
---|
134 | Yielder threads[nthreads]; |
---|
135 | printf("Starting\n"); |
---|
136 | |
---|
137 | bool is_tty = isatty(STDOUT_FILENO); |
---|
138 | start = getTimeNsec(); |
---|
139 | run = true; |
---|
140 | |
---|
141 | for(i; nthreads) { |
---|
142 | unpark( threads[i] __cfaabi_dbg_ctx2 ); |
---|
143 | } |
---|
144 | wait(duration, start, end, is_tty); |
---|
145 | |
---|
146 | run = false; |
---|
147 | end = getTimeNsec(); |
---|
148 | printf("\nDone\n"); |
---|
149 | } |
---|
150 | } |
---|
151 | |
---|
152 | printf("Took %'ld ms\n", (end - start)`ms); |
---|
153 | printf("Yields per second : %'18.2lf\n", ((double)global_counter) / (end - start)`s); |
---|
154 | printf("ns per yields : %'18.2lf\n", ((double)(end - start)`ns) / global_counter); |
---|
155 | printf("Total yields : %'15llu\n", global_counter); |
---|
156 | printf("Yields per procs : %'15llu\n", global_counter / nprocs); |
---|
157 | printf("Yields/sec/procs : %'18.2lf\n", (((double)global_counter) / nprocs) / (end - start)`s); |
---|
158 | printf("ns per yields/procs : %'18.2lf\n", ((double)(end - start)`ns) / (global_counter / nprocs)); |
---|
159 | |
---|
160 | } |
---|
161 | } |
---|