| 1 | #define _GNU_SOURCE
 | 
|---|
| 2 | 
 | 
|---|
| 3 | #include <stdlib.h>
 | 
|---|
| 4 | #include <stdio.h>
 | 
|---|
| 5 | #include <string.h>
 | 
|---|
| 6 | 
 | 
|---|
| 7 | extern "C" {
 | 
|---|
| 8 |         #include <locale.h>
 | 
|---|
| 9 |         #include <getopt.h>
 | 
|---|
| 10 |         #include <fcntl.h>
 | 
|---|
| 11 |         #include <sys/uio.h>
 | 
|---|
| 12 | }
 | 
|---|
| 13 | 
 | 
|---|
| 14 | #include <errno.h>
 | 
|---|
| 15 | #include <unistd.h>
 | 
|---|
| 16 | 
 | 
|---|
| 17 | #include <clock.hfa>
 | 
|---|
| 18 | #include <iofwd.hfa>
 | 
|---|
| 19 | #include <kernel.hfa>
 | 
|---|
| 20 | #include <thread.hfa>
 | 
|---|
| 21 | #include <time.hfa>
 | 
|---|
| 22 | #include <stats.hfa>
 | 
|---|
| 23 | 
 | 
|---|
| 24 | #include "../benchcltr.hfa"
 | 
|---|
| 25 | 
 | 
|---|
| 26 | extern bool traceHeapOn();
 | 
|---|
| 27 | 
 | 
|---|
| 28 | int fd;
 | 
|---|
| 29 | volatile bool run = false;
 | 
|---|
| 30 | volatile size_t count = 0;
 | 
|---|
| 31 | 
 | 
|---|
| 32 | unsigned long int buflen = 512;
 | 
|---|
| 33 | bool fixed_file = false;
 | 
|---|
| 34 | 
 | 
|---|
| 35 | thread __attribute__((aligned(128))) Reader {};
 | 
|---|
| 36 | void ?{}( Reader & this ) {
 | 
|---|
| 37 |         ((thread&)this){ "Reader Thread", *the_benchmark_cluster };
 | 
|---|
| 38 | }
 | 
|---|
| 39 | 
 | 
|---|
| 40 | int do_read(int fd, struct iovec * iov) {
 | 
|---|
| 41 |         // extern ssize_t cfa_preadv2(int, const struct iovec *, int, off_t, int, int = 0, Duration = -1`s, io_cancellation * = 0p, io_context * = 0p);
 | 
|---|
| 42 |         int sflags = 0
 | 
|---|
| 43 |         #if defined(CFA_HAVE_IOSQE_ASYNC)
 | 
|---|
| 44 |                 | CFA_IO_ASYNC
 | 
|---|
| 45 |         #else
 | 
|---|
| 46 |         #warning no CFA_IO_ASYNC support
 | 
|---|
| 47 |         #endif
 | 
|---|
| 48 |         ;
 | 
|---|
| 49 |         if(fixed_file) {
 | 
|---|
| 50 |                 sflags |= CFA_IO_FIXED_FD1;
 | 
|---|
| 51 |         }
 | 
|---|
| 52 |         return cfa_preadv2(fd, iov, 1, 0, 0, sflags, -1`s, 0p, 0p);
 | 
|---|
| 53 | }
 | 
|---|
| 54 | 
 | 
|---|
| 55 | void main( Reader & ) {
 | 
|---|
| 56 |         park();
 | 
|---|
| 57 |         /* paranoid */ assert( true == __atomic_load_n(&run, __ATOMIC_RELAXED) );
 | 
|---|
| 58 | 
 | 
|---|
| 59 |         __attribute__((aligned(512)))  char data[buflen];
 | 
|---|
| 60 |         struct iovec iov = { data, buflen };
 | 
|---|
| 61 | 
 | 
|---|
| 62 |         while(__atomic_load_n(&run, __ATOMIC_RELAXED)) {
 | 
|---|
| 63 |                 int r = do_read(fd, &iov);
 | 
|---|
| 64 |                 if(r < 0) abort("%s\n", strerror(errno));
 | 
|---|
| 65 | 
 | 
|---|
| 66 |                 __atomic_fetch_add( &count, 1, __ATOMIC_SEQ_CST );
 | 
|---|
| 67 |         }
 | 
|---|
| 68 | }
 | 
|---|
| 69 | 
 | 
|---|
| 70 | int main(int argc, char * argv[]) {
 | 
|---|
| 71 |         int file_flags = 0;
 | 
|---|
| 72 |         unsigned num_io = 1;
 | 
|---|
| 73 |         unsigned sublen = 16;
 | 
|---|
| 74 |         unsigned nentries = 0;
 | 
|---|
| 75 | 
 | 
|---|
| 76 |         bool subthrd = false;
 | 
|---|
| 77 |         bool subeagr = false;
 | 
|---|
| 78 |         bool odirect = false;
 | 
|---|
| 79 |         bool kpollsb = false;
 | 
|---|
| 80 |         bool kpollcp = false;
 | 
|---|
| 81 | 
 | 
|---|
| 82 |         cfa_option opt[] = {
 | 
|---|
| 83 |                 BENCH_OPT_CFA
 | 
|---|
| 84 |                 {'b', "bufsize",       "Number of bytes to read per request", buflen},
 | 
|---|
| 85 |                 {'s', "submitthread",  "If set, cluster uses polling thread to submit I/O", subthrd, parse_settrue},
 | 
|---|
| 86 |                 {'e', "eagersubmit",   "If set, cluster submits I/O eagerly but still aggregates submits", subeagr, parse_settrue},
 | 
|---|
| 87 |                 {'f', "fixed-files",   "Pre-register files with the io_contexts", fixed_file, parse_settrue},
 | 
|---|
| 88 |                 {'o', "open-direct",   "Open files with O_DIRECT flag, bypassing the file cache", odirect, parse_settrue},
 | 
|---|
| 89 |                 {'k', "kpollsubmit",   "If set, cluster uses an in kernel thread to poll submission, implies -f, requires elevated permissions", kpollsb, parse_settrue},
 | 
|---|
| 90 |                 {'i', "kpollcomplete", "If set, cluster polls fds for completions instead of relying on interrupts to get notifications, implies -o", kpollcp, parse_settrue},
 | 
|---|
| 91 |                 {'l', "submitlength",  "Size of the buffer that stores ready submissions", sublen},
 | 
|---|
| 92 |                 {'r', "numentries",    "Number of entries each of the io_context have", nentries},
 | 
|---|
| 93 |                 {'n', "numcontexts",   "Number of io_contexts to the cluster", num_io},
 | 
|---|
| 94 |         };
 | 
|---|
| 95 |         int opt_cnt = sizeof(opt) / sizeof(cfa_option);
 | 
|---|
| 96 | 
 | 
|---|
| 97 |         char **left;
 | 
|---|
| 98 |         parse_args( opt, opt_cnt, "[OPTIONS]...\ncforall readv benchmark", left );
 | 
|---|
| 99 | 
 | 
|---|
| 100 |         if(kpollcp || odirect) {
 | 
|---|
| 101 |                 if( (buflen % 512) != 0 ) {
 | 
|---|
| 102 |                         fprintf(stderr, "Buffer length must be a multiple of 512 when using O_DIRECT, was %lu\n\n", buflen);
 | 
|---|
| 103 |                         print_args_usage(opt, opt_cnt, "[OPTIONS]...\ncforall readv benchmark", true);
 | 
|---|
| 104 |                 }
 | 
|---|
| 105 |         }
 | 
|---|
| 106 | 
 | 
|---|
| 107 |         io_context_params params;
 | 
|---|
| 108 | 
 | 
|---|
| 109 |         if( subthrd ) params.poller_submits = true;
 | 
|---|
| 110 |         if( subeagr ) params.eager_submits  = true;
 | 
|---|
| 111 |         if( kpollsb ) params.poll_submit    = true;
 | 
|---|
| 112 |         if( kpollcp ) params.poll_complete  = true;
 | 
|---|
| 113 | 
 | 
|---|
| 114 |         if(params.poll_submit  ) fixed_file = true;
 | 
|---|
| 115 |         if(params.poll_complete) odirect    = true;
 | 
|---|
| 116 | 
 | 
|---|
| 117 |         params.num_ready = sublen;
 | 
|---|
| 118 |         params.num_entries = nentries;
 | 
|---|
| 119 | 
 | 
|---|
| 120 |         if(odirect) file_flags |= O_DIRECT;
 | 
|---|
| 121 | 
 | 
|---|
| 122 |         int lfd = open(__FILE__, file_flags);
 | 
|---|
| 123 |         if(lfd < 0) {
 | 
|---|
| 124 |                 fprintf(stderr, "Could not open source file\n");
 | 
|---|
| 125 |                 exit(EXIT_FAILURE);
 | 
|---|
| 126 |         }
 | 
|---|
| 127 | 
 | 
|---|
| 128 |         printf("Running %d threads, reading %lu bytes each, over %d processors for %f seconds\n", nthreads, buflen, nprocs, duration);
 | 
|---|
| 129 | 
 | 
|---|
| 130 |         {
 | 
|---|
| 131 |                 Time start, end;
 | 
|---|
| 132 |                 BenchCluster cl = { num_io, params, CFA_STATS_READY_Q | CFA_STATS_IO };
 | 
|---|
| 133 | 
 | 
|---|
| 134 |                 if(fixed_file) {
 | 
|---|
| 135 |                         fd = 0;
 | 
|---|
| 136 |                         register_fixed_files( cl.self, &lfd, 1 );
 | 
|---|
| 137 |                 }
 | 
|---|
| 138 |                 else {
 | 
|---|
| 139 |                         fd = lfd;
 | 
|---|
| 140 |                 }
 | 
|---|
| 141 | 
 | 
|---|
| 142 |                 {
 | 
|---|
| 143 |                         BenchProc procs[nprocs];
 | 
|---|
| 144 |                         {
 | 
|---|
| 145 |                                 Reader threads[nthreads];
 | 
|---|
| 146 | 
 | 
|---|
| 147 |                                 printf("Starting\n");
 | 
|---|
| 148 |                                 bool is_tty = isatty(STDOUT_FILENO);
 | 
|---|
| 149 |                                 start = getTimeNsec();
 | 
|---|
| 150 |                                 run = true;
 | 
|---|
| 151 | 
 | 
|---|
| 152 |                                 for(i; nthreads) {
 | 
|---|
| 153 |                                         unpark( threads[i] );
 | 
|---|
| 154 |                                 }
 | 
|---|
| 155 |                                 wait(duration, start, end, is_tty);
 | 
|---|
| 156 | 
 | 
|---|
| 157 |                                 run = false;
 | 
|---|
| 158 |                                 end = getTimeNsec();
 | 
|---|
| 159 |                                 printf("\nDone\n");
 | 
|---|
| 160 |                         }
 | 
|---|
| 161 |                         printf("Readers closed\n");
 | 
|---|
| 162 |                 }
 | 
|---|
| 163 |                 printf("Took %'ld ms\n", (end - start)`ms);
 | 
|---|
| 164 |                 printf("Total reads      : %'15zu\n", count);
 | 
|---|
| 165 |                 printf("Reads per second : %'18.2lf\n", ((double)count) / (end - start)`s);
 | 
|---|
| 166 |                 printf("Total read size  : %'15zu\n", buflen * count);
 | 
|---|
| 167 |                 printf("Bytes per second : %'18.2lf\n", ((double)count * buflen) / (end - start)`s);
 | 
|---|
| 168 |         }
 | 
|---|
| 169 | 
 | 
|---|
| 170 |         close(lfd);
 | 
|---|
| 171 | }
 | 
|---|