1 | #include <stdlib.h> |
---|
2 | #include <stdio.h> |
---|
3 | #include <string.h> |
---|
4 | |
---|
5 | extern "C" { |
---|
6 | #include <locale.h> |
---|
7 | #include <getopt.h> |
---|
8 | #include <fcntl.h> |
---|
9 | #include <sys/uio.h> |
---|
10 | } |
---|
11 | |
---|
12 | #include <unistd.h> |
---|
13 | |
---|
14 | #include <clock.hfa> |
---|
15 | #include <kernel.hfa> |
---|
16 | #include <thread.hfa> |
---|
17 | #include <time.hfa> |
---|
18 | #include <stats.hfa> |
---|
19 | |
---|
20 | #include "../benchcltr.hfa" |
---|
21 | |
---|
22 | extern bool traceHeapOn(); |
---|
23 | extern ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags); |
---|
24 | |
---|
25 | int fd; |
---|
26 | volatile bool run = false; |
---|
27 | volatile size_t count = 0; |
---|
28 | |
---|
29 | unsigned long int buflen = 50; |
---|
30 | |
---|
31 | thread __attribute__((aligned(128))) Reader {}; |
---|
32 | void ?{}( Reader & this ) { |
---|
33 | ((thread&)this){ "Reader Thread", *the_benchmark_cluster }; |
---|
34 | } |
---|
35 | |
---|
36 | void main( Reader & ) { |
---|
37 | park( __cfaabi_dbg_ctx ); |
---|
38 | /* paranoid */ assert( true == __atomic_load_n(&run, __ATOMIC_RELAXED) ); |
---|
39 | |
---|
40 | char data[buflen]; |
---|
41 | struct iovec iov = { data, buflen }; |
---|
42 | |
---|
43 | while(__atomic_load_n(&run, __ATOMIC_RELAXED)) { |
---|
44 | int r = cfa_preadv2(fd, &iov, 1, 0, 0); |
---|
45 | if(r < 0) abort("%s\n", strerror(-r)); |
---|
46 | |
---|
47 | __atomic_fetch_add( &count, 1, __ATOMIC_SEQ_CST ); |
---|
48 | } |
---|
49 | } |
---|
50 | |
---|
51 | int main(int argc, char * argv[]) { |
---|
52 | BENCH_DECL |
---|
53 | unsigned flags = 0; |
---|
54 | unsigned sublen = 16; |
---|
55 | |
---|
56 | arg_loop: |
---|
57 | for(;;) { |
---|
58 | static struct option options[] = { |
---|
59 | BENCH_OPT_LONG |
---|
60 | {"bufsize", required_argument, 0, 'b'}, |
---|
61 | {"userthread", no_argument , 0, 'u'}, |
---|
62 | {"submitthread", no_argument , 0, 's'}, |
---|
63 | {"eagersubmit", no_argument , 0, 'e'}, |
---|
64 | {"kpollsubmit", no_argument , 0, 'k'}, |
---|
65 | {"kpollcomplete", no_argument , 0, 'i'}, |
---|
66 | {"submitlength", required_argument, 0, 'l'}, |
---|
67 | {0, 0, 0, 0} |
---|
68 | }; |
---|
69 | |
---|
70 | int idx = 0; |
---|
71 | int opt = getopt_long(argc, argv, BENCH_OPT_SHORT "b:usekil:", options, &idx); |
---|
72 | |
---|
73 | const char * arg = optarg ? optarg : ""; |
---|
74 | char * end; |
---|
75 | switch(opt) { |
---|
76 | // Exit Case |
---|
77 | case -1: |
---|
78 | break arg_loop; |
---|
79 | BENCH_OPT_CASE |
---|
80 | case 'b': |
---|
81 | buflen = strtoul(arg, &end, 10); |
---|
82 | if(*end != '\0' && buflen < 10) { |
---|
83 | fprintf(stderr, "Buffer size must be at least 10, was %s\n", arg); |
---|
84 | goto usage; |
---|
85 | } |
---|
86 | break; |
---|
87 | case 'u': |
---|
88 | flags |= CFA_CLUSTER_IO_POLLER_USER_THREAD; |
---|
89 | break; |
---|
90 | case 's': |
---|
91 | flags |= CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS; |
---|
92 | break; |
---|
93 | case 'e': |
---|
94 | flags |= CFA_CLUSTER_IO_EAGER_SUBMITS; |
---|
95 | break; |
---|
96 | case 'k': |
---|
97 | flags |= CFA_CLUSTER_IO_KERNEL_POLL_SUBMITS; |
---|
98 | break; |
---|
99 | case 'i': |
---|
100 | flags |= CFA_CLUSTER_IO_KERNEL_POLL_COMPLETES; |
---|
101 | break; |
---|
102 | case 'l': |
---|
103 | sublen = strtoul(arg, &end, 10); |
---|
104 | if(*end != '\0' && sublen < 16) { |
---|
105 | fprintf(stderr, "Submit length must be at least 16, was %s\n", arg); |
---|
106 | goto usage; |
---|
107 | } |
---|
108 | flags |= (sublen << CFA_CLUSTER_IO_BUFFLEN_OFFSET); |
---|
109 | break; |
---|
110 | default: /* ? */ |
---|
111 | fprintf(stderr, "%d\n", opt); |
---|
112 | usage: |
---|
113 | bench_usage( argv ); |
---|
114 | fprintf( stderr, " -b, --buflen=SIZE Number of bytes to read per request\n" ); |
---|
115 | fprintf( stderr, " -u, --userthread If set, cluster uses user-thread to poll I/O\n" ); |
---|
116 | fprintf( stderr, " -s, --submitthread If set, cluster uses polling thread to submit I/O\n" ); |
---|
117 | fprintf( stderr, " -e, --eagersubmit If set, cluster submits I/O eagerly but still aggregates submits\n" ); |
---|
118 | fprintf( stderr, " -k, --kpollsubmit If set, cluster uses IORING_SETUP_SQPOLL\n" ); |
---|
119 | fprintf( stderr, " -i, --kpollcomplete If set, cluster uses IORING_SETUP_IOPOLL\n" ); |
---|
120 | fprintf( stderr, " -l, --submitlength=LEN Max number of submitions that can be submitted together\n" ); |
---|
121 | exit(EXIT_FAILURE); |
---|
122 | } |
---|
123 | } |
---|
124 | |
---|
125 | fd = open(__FILE__, 0); |
---|
126 | if(fd < 0) { |
---|
127 | fprintf(stderr, "Could not open source file\n"); |
---|
128 | exit(EXIT_FAILURE); |
---|
129 | } |
---|
130 | |
---|
131 | printf("Running %d threads, reading %lu bytes each, over %d processors for %f seconds\n", nthreads, buflen, nprocs, duration); |
---|
132 | |
---|
133 | { |
---|
134 | Time start, end; |
---|
135 | BenchCluster cl = { flags, CFA_STATS_READY_Q | CFA_STATS_IO }; |
---|
136 | { |
---|
137 | BenchProc procs[nprocs]; |
---|
138 | { |
---|
139 | Reader threads[nthreads]; |
---|
140 | |
---|
141 | printf("Starting\n"); |
---|
142 | bool is_tty = isatty(STDOUT_FILENO); |
---|
143 | start = getTimeNsec(); |
---|
144 | run = true; |
---|
145 | |
---|
146 | for(i; nthreads) { |
---|
147 | unpark( threads[i] __cfaabi_dbg_ctx2 ); |
---|
148 | } |
---|
149 | wait(duration, start, end, is_tty); |
---|
150 | |
---|
151 | run = false; |
---|
152 | end = getTimeNsec(); |
---|
153 | printf("\nDone\n"); |
---|
154 | } |
---|
155 | } |
---|
156 | printf("Took %'ld ms\n", (end - start)`ms); |
---|
157 | printf("Total reads : %'15zu\n", count); |
---|
158 | printf("Reads per second : %'18.2lf\n", ((double)count) / (end - start)`s); |
---|
159 | printf("Total read size : %'15zu\n", buflen * count); |
---|
160 | printf("Bytes per second : %'18.2lf\n", ((double)count * buflen) / (end - start)`s); |
---|
161 | } |
---|
162 | |
---|
163 | close(fd); |
---|
164 | } |
---|