| 1 | #pragma once
|
|---|
| 2 | #include <assert.h>
|
|---|
| 3 | #include <stdint.h>
|
|---|
| 4 |
|
|---|
| 5 | #ifdef __cforall
|
|---|
| 6 | #include <kernel.hfa>
|
|---|
| 7 | #include <thread.hfa>
|
|---|
| 8 | #include <stats.hfa>
|
|---|
| 9 | #else
|
|---|
| 10 | #include <time.h> // timespec
|
|---|
| 11 | #include <sys/time.h> // timeval
|
|---|
| 12 |
|
|---|
| 13 | enum { TIMEGRAN = 1000000000LL }; // nanosecond granularity, except for timeval
|
|---|
| 14 | #endif
|
|---|
| 15 |
|
|---|
| 16 | #define BENCH_OPT_SHORT "d:p:t:SPV"
|
|---|
| 17 | #define BENCH_OPT_LONG \
|
|---|
| 18 | {"duration", required_argument, 0, 'd'}, \
|
|---|
| 19 | {"nthreads", required_argument, 0, 't'}, \
|
|---|
| 20 | {"nprocs", required_argument, 0, 'p'}, \
|
|---|
| 21 | {"nostats", no_argument , 0, 'S'}, \
|
|---|
| 22 | {"procstat", no_argument , 0, 'P'}, \
|
|---|
| 23 | {"viewhalts", no_argument , 0, 'V'},
|
|---|
| 24 |
|
|---|
| 25 | #define BENCH_DECL \
|
|---|
| 26 | double duration = 5; \
|
|---|
| 27 | int nprocs = 1; \
|
|---|
| 28 | int nthreads = 1;
|
|---|
| 29 |
|
|---|
| 30 | #define BENCH_OPT_CASE \
|
|---|
| 31 | case 'd': \
|
|---|
| 32 | duration = strtod(arg, &end); \
|
|---|
| 33 | if(*end != '\0') { \
|
|---|
| 34 | fprintf(stderr, "Duration must be a valid double, was %s\n", arg); \
|
|---|
| 35 | goto usage; \
|
|---|
| 36 | } \
|
|---|
| 37 | break; \
|
|---|
| 38 | case 't': \
|
|---|
| 39 | nthreads = strtoul(arg, &end, 10); \
|
|---|
| 40 | if(*end != '\0' || nthreads < 1) { \
|
|---|
| 41 | fprintf(stderr, "Number of threads must be a positive integer, was %s\n", arg); \
|
|---|
| 42 | goto usage; \
|
|---|
| 43 | } \
|
|---|
| 44 | break; \
|
|---|
| 45 | case 'p': \
|
|---|
| 46 | nprocs = strtoul(arg, &end, 10); \
|
|---|
| 47 | if(*end != '\0' || nprocs < 1) { \
|
|---|
| 48 | fprintf(stderr, "Number of processors must be a positive integer, was %s\n", arg); \
|
|---|
| 49 | goto usage; \
|
|---|
| 50 | } \
|
|---|
| 51 | break; \
|
|---|
| 52 | case 'S': \
|
|---|
| 53 | silent = true; \
|
|---|
| 54 | break; \
|
|---|
| 55 | case 'P': \
|
|---|
| 56 | procstats = true; \
|
|---|
| 57 | break; \
|
|---|
| 58 | case 'V': \
|
|---|
| 59 | viewhalts = true; \
|
|---|
| 60 | break;
|
|---|
| 61 |
|
|---|
| 62 | bool silent = false;
|
|---|
| 63 | bool procstats = false;
|
|---|
| 64 | bool viewhalts = false;
|
|---|
| 65 |
|
|---|
| 66 | #ifdef __cforall
|
|---|
| 67 | struct cluster * the_benchmark_cluster = 0p;
|
|---|
| 68 | struct BenchCluster {
|
|---|
| 69 | cluster self;
|
|---|
| 70 | };
|
|---|
| 71 |
|
|---|
| 72 | void ?{}( BenchCluster & this, int num_io, const io_context_params & io_params, int stats ) {
|
|---|
| 73 | (this.self){ "Benchmark Cluster", num_io, io_params };
|
|---|
| 74 |
|
|---|
| 75 | assert( the_benchmark_cluster == 0p );
|
|---|
| 76 | the_benchmark_cluster = &this.self;
|
|---|
| 77 |
|
|---|
| 78 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 79 | if( !silent ) {
|
|---|
| 80 | print_stats_at_exit( this.self, stats );
|
|---|
| 81 | }
|
|---|
| 82 | #endif
|
|---|
| 83 | }
|
|---|
| 84 |
|
|---|
| 85 | struct BenchProc {
|
|---|
| 86 | processor self;
|
|---|
| 87 | };
|
|---|
| 88 |
|
|---|
| 89 | void ?{}( BenchProc & this ) {
|
|---|
| 90 | assert( the_benchmark_cluster != 0p );
|
|---|
| 91 | (this.self){ "Benchmark Processor", *the_benchmark_cluster };
|
|---|
| 92 |
|
|---|
| 93 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 94 | if( procstats ) {
|
|---|
| 95 | print_stats_at_exit( this.self, the_benchmark_cluster->print_stats );
|
|---|
| 96 | }
|
|---|
| 97 | if( viewhalts ) {
|
|---|
| 98 | print_halts( this.self );
|
|---|
| 99 | }
|
|---|
| 100 | #endif
|
|---|
| 101 | }
|
|---|
| 102 |
|
|---|
| 103 | void wait(double duration, Time & start, Time & end, bool is_tty) {
|
|---|
| 104 | for() {
|
|---|
| 105 | sleep(100`ms);
|
|---|
| 106 | end = getTimeNsec();
|
|---|
| 107 | Duration delta = end - start;
|
|---|
| 108 | /*if(is_tty)*/ {
|
|---|
| 109 | printf(" %.1f\r", delta`ds);
|
|---|
| 110 | fflush(stdout);
|
|---|
| 111 | }
|
|---|
| 112 | if( delta >= duration`s ) {
|
|---|
| 113 | break;
|
|---|
| 114 | }
|
|---|
| 115 | }
|
|---|
| 116 | }
|
|---|
| 117 | #else
|
|---|
| 118 | uint64_t getTimeNsec() {
|
|---|
| 119 | timespec curr;
|
|---|
| 120 | clock_gettime( CLOCK_REALTIME, &curr );
|
|---|
| 121 | return (int64_t)curr.tv_sec * TIMEGRAN + curr.tv_nsec;
|
|---|
| 122 | }
|
|---|
| 123 |
|
|---|
| 124 | uint64_t to_miliseconds( uint64_t durtn ) { return durtn / (TIMEGRAN / 1000LL); }
|
|---|
| 125 | double to_fseconds(uint64_t durtn ) { return durtn / (double)TIMEGRAN; }
|
|---|
| 126 | uint64_t from_fseconds(double sec) { return sec * TIMEGRAN; }
|
|---|
| 127 |
|
|---|
| 128 |
|
|---|
| 129 | void wait_duration(double duration, uint64_t & start, uint64_t & end, bool is_tty) {
|
|---|
| 130 | for(;;) {
|
|---|
| 131 | usleep(100000);
|
|---|
| 132 | end = getTimeNsec();
|
|---|
| 133 | uint64_t delta = end - start;
|
|---|
| 134 | /*if(is_tty)*/ {
|
|---|
| 135 | printf(" %.1f\r", to_fseconds(delta));
|
|---|
| 136 | fflush(stdout);
|
|---|
| 137 | }
|
|---|
| 138 | if( delta >= from_fseconds(duration) ) {
|
|---|
| 139 | break;
|
|---|
| 140 | }
|
|---|
| 141 | }
|
|---|
| 142 | }
|
|---|
| 143 | #endif
|
|---|
| 144 |
|
|---|
| 145 |
|
|---|
| 146 | void bench_usage( char * argv [] ) {
|
|---|
| 147 | fprintf( stderr, "Usage: %s : [options]\n", argv[0] );
|
|---|
| 148 | fprintf( stderr, "\n" );
|
|---|
| 149 | fprintf( stderr, " -d, --duration=DURATION Duration of the experiment, in seconds\n" );
|
|---|
| 150 | fprintf( stderr, " -t, --nthreads=NTHREADS Number of user threads\n" );
|
|---|
| 151 | fprintf( stderr, " -p, --nprocs=NPROCS Number of kernel threads\n" );
|
|---|
| 152 | fprintf( stderr, " -S, --nostats Don't print cluster stats\n" );
|
|---|
| 153 | fprintf( stderr, " -P, --procstat Print processor stats" );
|
|---|
| 154 | }
|
|---|