[splice.git] / splice-test4c.c

#include <assert.h>
#include <ctype.h>
#include <errno.h>
#include <fcntl.h>
#include <malloc.h>
#include <netdb.h>
#include <netinet/in.h>
#include <netinet/tcp.h>
#include <sched.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <sys/sendfile.h>
#include <sys/socket.h>
#include <sys/time.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <time.h>
#include <unistd.h>

#include "splice.h"

#define TARGET_HOSTNAME "localhost"

#define BYTES (128*1024*1024UL)
#define BUFSIZE (64*1024)

#define NR (BYTES/BUFSIZE)

#define SENDFILE_LOOPS 10
#define SPLICE_LOOPS 10
#define SPLICE_PIPE_LOOPS 10

static int sendfile_loops = SENDFILE_LOOPS;
static int splice_pipe_loops = SPLICE_PIPE_LOOPS;
#if 0
static int splice_loops = SPLICE_LOOPS;
#endif

static volatile long long *cycles, cycles_per_sec;

static struct timeval start_time;
static double start_cycles;
static double cpu_pct;

static void start_timing(const char *desc)
{
	printf("%-20s: ", desc);
	fflush(stdout);
	gettimeofday(&start_time, NULL);
	/*
	 * Give the lowprio cycles thread a chance to run and thus
	 * we get an accurate timestamp:
	 */
	sched_yield();
	start_cycles = (double)*cycles;
}

static double end_timing(unsigned long long bytes, double *rate)
{
	static long long total;
	struct timeval end_time;
	double usecs;
	double end_cycles, cpu_cycles;

	gettimeofday(&end_time, NULL);
	end_cycles = (double)*cycles;

	usecs = (double) (end_time.tv_sec - start_time.tv_sec);
	usecs *= 1000000.0;
	usecs += (double) (end_time.tv_usec - start_time.tv_usec);
	total += bytes;

	cpu_cycles = end_cycles - start_cycles;
	cpu_pct = 100.0 -
		cpu_cycles / cycles_per_sec / ( usecs / 1000000.0 ) * 100.0;

	*rate = (double) bytes / usecs / (1024*1024) * 1000000;

	printf("%.2fMB/s (%.1fMB total, %.2f%% CPU)\n", *rate,
		(double) total / (1024*1024),
		cpu_pct
	);

	return cpu_pct;
}

static void calibrate_loops(void)
{
	long long l0, l1;
	int i;

	cycles_per_sec = 0;
	printf("calibrating cycles: "); fflush(stdout);

	/*
	 * Make sure we start on a precise timer IRQ boundary:
	 */
	usleep(50000);

	for (i = 0; i < 10; i++) {
		sched_yield();
		l0 = *cycles;
		usleep(200000);
		l1 = *cycles;
		cycles_per_sec = max(cycles_per_sec, l1-l0);
	}
	cycles_per_sec *= 5;

	printf("%Ld cycles/sec\n", cycles_per_sec);
}

static int child(struct sockaddr *addr, int len)
{
	static char buffer[BUFSIZE];
	int sk;
	double c1, c2, c3;
	int fd;
	struct sockaddr_in s_to;
	struct hostent *hp;
	double r1, r2, r3, r4, r5;
	int i, pipefd[2];
	loff_t off = 0;

	r1 = r2 = r3 = r4 = r5 = 0;

	sk = socket(PF_INET, SOCK_STREAM, 0);
	if (!sk)
		return error("socket");
	hp = gethostbyname (TARGET_HOSTNAME);
	BUG_ON(!hp);
	bzero ((char *) &s_to, sizeof (s_to));
	bcopy ((char *) hp->h_addr, (char *) &(s_to.sin_addr), hp->h_length);
	s_to.sin_family = hp->h_addrtype;
	s_to.sin_port = htons(1111);

	calibrate_loops();

	fprintf(stdout, "BUFSIZE = %d\n", BUFSIZE);
	fflush(stdout);

	if (connect(sk, (struct sockaddr *)&s_to, len) < 0)
		return error("connect");

	start_timing("Empty buffer");
	for (i = 0; i < NR; i++)
		write(sk, buffer, BUFSIZE);
	end_timing(NR*BUFSIZE, &r1);

	fd = open("largefile", O_RDONLY);
	if (fd < 0)
		return error("largefile");

	start_timing("Read/write loop");
	for (i = 0; i < NR; i++) {
		if (read(fd, buffer, BUFSIZE) != BUFSIZE)
			return error("largefile read");
		write(sk, buffer, BUFSIZE);
	}
	end_timing(NR*BUFSIZE, &r2);
	close(fd);
	close(sk);

	start_timing("sendfile");
sendfile_again:
	sk = socket(PF_INET, SOCK_STREAM, 0);
	if (connect(sk, (struct sockaddr *)&s_to, len) < 0)
		return error("connect");

	fd = open("largefile", O_RDONLY);
	if (fd < 0)
		return error("largefile");

	i = NR*BUFSIZE;
	do {
		int ret = sendfile(sk, fd, NULL, i);
		i -= ret;
	} while (i);

	close(fd);
	close(sk);
	if (--sendfile_loops)
		goto sendfile_again;
	c1 = end_timing(NR*BUFSIZE*SENDFILE_LOOPS, &r3);

	start_timing("splice-pipe");
splice_pipe_again:
	sk = socket(PF_INET, SOCK_STREAM, 0);
	if (connect(sk, (struct sockaddr *)&s_to, len) < 0)
		return error("connect");

	fd = open("largefile", O_RDONLY);
	if (fd < 0)
		return error("largefile");
	if (pipe(pipefd) < 0)
		return error("pipe");

	i = NR*BUFSIZE;
	off = 0;
	do {
		int ret = splice(fd, &off, pipefd[1], NULL, min(i, BUFSIZE), SPLICE_F_NONBLOCK);
		if (ret <= 0)
			return error("splice-pipe-in");
		i -= ret;
		while (ret > 0) {
			int flags = i ? SPLICE_F_MORE : 0;
			int written = splice(pipefd[0], NULL, sk, NULL, ret, flags);
			if (written <= 0)
				return error("splice-pipe-out");
			ret -= written;
		}
	} while (i);

	close(fd);
	close(sk);
	close(pipefd[0]);
	close(pipefd[1]);
	if (--splice_pipe_loops)
		goto splice_pipe_again;
	c2 = end_timing(NR*BUFSIZE*SPLICE_LOOPS, &r4);

	/*
	 * Direct splicing was disabled as being immediately available,
	 * it's reserved for sendfile emulation now.
	 */
#if 0
	start_timing("splice");
splice_again:
	sk = socket(PF_INET, SOCK_STREAM, 0);
	if (connect(sk, (struct sockaddr *)&s_to, len) < 0)
		return error("connect");

	fd = open("largefile", O_RDONLY);
	if (fd < 0)
		return error("largefile");

	i = NR*BUFSIZE;
	off = 0;
	do {
		int flags = BUFSIZE < i ? SPLICE_F_MORE : 0;
		int ret;

		ret = splice(fd, &off, sk, NULL, min(i, BUFSIZE), flags);

		if (ret <= 0)
			return error("splice");
		i -= ret;
	} while (i);

	close(fd);
	close(sk);
	if (--splice_loops)
		goto splice_again;
	c3 = end_timing(NR*BUFSIZE*SPLICE_LOOPS, &r5);
#else
	c3 = 0;
#endif

	/*
	 * c1/r3 - sendfile
	 * c2/r4 - splice-pipe
	 * c3/r5 - splice
	 */

	if (c1 && c2)
		printf("sendfile is %.2f%% more efficient than splice-pipe.\n",
			(c2 - c1) / c1 * 100.0 );
	if (c1 && c3)
		printf("sendfile is %.2f%% more efficient than splice.\n",
			(c3 - c1) / c1 * 100.0 );
	if (c2 && c3)
		printf("splice is %.2f%% more efficient splice-pipe.\n",
			(c2 - c3) / c3 * 100.0 );
	if (r3 && r4)
		printf("sendfile is %.2f%% faster than splice-pipe.\n",
			(r3 - r4) / r4 * 100.0 );
	if (r3 && r5)
		printf("sendfile is %.2f%% faster than splice.\n",
			(r3 - r5) / r5 * 100.0 );
	if (r4 && r5)
		printf("splice is %.2f%% faster than splice-pipe.\n",
			(r5 - r4) / r4 * 100.0 );

	return 0;
}


static void setup_shared_var(void)
{
	char zerobuff [4096] = { 0, };
	int ret, fd;

	fd = creat(".tmp_mmap", 0700);
	BUG_ON(fd == -1);
	close(fd);

	fd = open(".tmp_mmap", O_RDWR|O_CREAT|O_TRUNC);
	BUG_ON(fd == -1);
	ret = write(fd, zerobuff, 4096);
	BUG_ON(ret != 4096);

	cycles = (void *)mmap(0, 4096, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
	BUG_ON(cycles == (void *)-1);

	close(fd);
}

#define SCHED_BATCH 3

#if defined(__i386__)
#define rdtscll(val)					\
do {							\
	__asm__ __volatile__("rdtsc" : "=A" (val));	\
} while (0)
#if 0
#elif defined(__ia64__)
#define rdtscll(val)					\
do {							\
	val = *__mm_clock_dev;				\
} while (0)
#endif
#else
#define rdtscll(val) \
	do { (val) = 0LL; } while (0)
#endif

/*
 * Keep lowprio looping - to meausure the number of idle cycles
 * available. It's tricky: we do a series of RDTSC calls, and
 * if the delay to the last measurement was less than 500 cycles,
 * we conclude that only this loop ran.
 */
static void lowprio_cycle_soak_loop(void)
{
        struct sched_param p = { sched_priority: 0 };
	unsigned long long t0, t1, delta;

	/*
	 * We are a nice +19 SCHED_BATCH task:
	 */
	BUG_ON(sched_setscheduler(0, SCHED_BATCH, &p) != 0);
	nice(40);

	rdtscll(t0);
	while (cycles >= 0) {
		rdtscll(t1);
		delta = t1-t0;
		if (delta < 500)
			*cycles += delta;
		t0 = t1;
	}
}

int main(int argc, char **argv)
{
	unsigned int sk, len;
	struct sockaddr addr;
	pid_t pid;

	setup_shared_var();

	signal(SIGCHLD, SIG_IGN);
	sk = socket(PF_INET, SOCK_STREAM, 0);
	if (sk < 0) {
		perror("socket");
		exit(1);
	}
	if (listen(sk, 1) < 0) {
		perror("listen");
		exit(1);
	}
	len = sizeof(addr);
	if (getsockname(sk, &addr, &len) < 0) {
		perror("getsockname");
		exit(1);
	}
	pid = fork();
	if (!pid) {
		lowprio_cycle_soak_loop();
		exit(0);
	}
	nice(-20);
	child(&addr, len);
	kill(pid, SIGHUP);
	exit(0);
}
Commit	Line	Data
d8525fbd JA	1	#include <assert.h>
	2	#include <ctype.h>
	3	#include <errno.h>
	4	#include <fcntl.h>
	5	#include <malloc.h>
	6	#include <netdb.h>
	7	#include <netinet/in.h>
	8	#include <netinet/tcp.h>
	9	#include <sched.h>
	10	#include <signal.h>
	11	#include <stdio.h>
	12	#include <stdlib.h>
	13	#include <string.h>
	14	#include <sys/mman.h>
	15	#include <sys/sendfile.h>
	16	#include <sys/socket.h>
	17	#include <sys/time.h>
	18	#include <sys/types.h>
	19	#include <sys/wait.h>
	20	#include <time.h>
	21	#include <unistd.h>
	22
	23	#include "splice.h"
	24
	25	#define TARGET_HOSTNAME "localhost"
	26
	27	#define BYTES (12810241024UL)
	28	#define BUFSIZE (64*1024)
	29
	30	#define NR (BYTES/BUFSIZE)
	31
	32	#define SENDFILE_LOOPS 10
	33	#define SPLICE_LOOPS 10
	34	#define SPLICE_PIPE_LOOPS 10
	35
	36	static int sendfile_loops = SENDFILE_LOOPS;
d8525fbd	37	static int splice_pipe_loops = SPLICE_PIPE_LOOPS;
422d15ba JA	38	#if 0
	39	static int splice_loops = SPLICE_LOOPS;
	40	#endif
d8525fbd JA	41
	42	static volatile long long *cycles, cycles_per_sec;
	43
	44	static struct timeval start_time;
	45	static double start_cycles;
	46	static double cpu_pct;
	47
	48	static void start_timing(const char *desc)
	49	{
	50	printf("%-20s: ", desc);
	51	fflush(stdout);
	52	gettimeofday(&start_time, NULL);
	53	/*
	54	* Give the lowprio cycles thread a chance to run and thus
	55	* we get an accurate timestamp:
	56	*/
	57	sched_yield();
	58	start_cycles = (double)*cycles;
	59	}
	60
	61	static double end_timing(unsigned long long bytes, double *rate)
	62	{
	63	static long long total;
	64	struct timeval end_time;
	65	double usecs;
	66	double end_cycles, cpu_cycles;
	67
	68	gettimeofday(&end_time, NULL);
	69	end_cycles = (double)*cycles;
	70
	71	usecs = (double) (end_time.tv_sec - start_time.tv_sec);
	72	usecs *= 1000000.0;
	73	usecs += (double) (end_time.tv_usec - start_time.tv_usec);
	74	total += bytes;
	75
	76	cpu_cycles = end_cycles - start_cycles;
	77	cpu_pct = 100.0 -
	78	cpu_cycles / cycles_per_sec / ( usecs / 1000000.0 ) * 100.0;
	79
	80	rate = (double) bytes / usecs / (10241024) * 1000000;
	81
	82	printf("%.2fMB/s (%.1fMB total, %.2f%% CPU)\n", *rate,
	83	(double) total / (1024*1024),
	84	cpu_pct
	85	);
	86
	87	return cpu_pct;
	88	}
	89
	90	static void calibrate_loops(void)
	91	{
	92	long long l0, l1;
	93	int i;
	94
	95	cycles_per_sec = 0;
	96	printf("calibrating cycles: "); fflush(stdout);
	97
	98	/*
	99	* Make sure we start on a precise timer IRQ boundary:
	100	*/
	101	usleep(50000);
	102
	103	for (i = 0; i < 10; i++) {
	104	sched_yield();
105	l0 = *cycles;
106	usleep(200000);
107	l1 = *cycles;
108	cycles_per_sec = max(cycles_per_sec, l1-l0);
109	}
110	cycles_per_sec *= 5;
111
112	printf("%Ld cycles/sec\n", cycles_per_sec);
113	}
114
115	static int child(struct sockaddr *addr, int len)
116	{
117	static char buffer[BUFSIZE];
118	int sk;
119	double c1, c2, c3;
120	int fd;
121	struct sockaddr_in s_to;
122	struct hostent *hp;
123	double r1, r2, r3, r4, r5;
124	int i, pipefd[2];
125	loff_t off = 0;
126
127	r1 = r2 = r3 = r4 = r5 = 0;
128
129	sk = socket(PF_INET, SOCK_STREAM, 0);
130	if (!sk)
131	return error("socket");
132	hp = gethostbyname (TARGET_HOSTNAME);
133	BUG_ON(!hp);
134	bzero ((char *) &s_to, sizeof (s_to));
135	bcopy ((char ) hp->h_addr, (char ) &(s_to.sin_addr), hp->h_length);
136	s_to.sin_family = hp->h_addrtype;
137	s_to.sin_port = htons(1111);
138
139	calibrate_loops();
140
141	fprintf(stdout, "BUFSIZE = %d\n", BUFSIZE);
142	fflush(stdout);
143
422d15ba JA	144	if (connect(sk, (struct sockaddr *)&s_to, len) < 0)
422d15ba JA	145	return error("connect");
d8525fbd	146
422d15ba JA	147	start_timing("Empty buffer");
	148	for (i = 0; i < NR; i++)
	149	write(sk, buffer, BUFSIZE);
	150	end_timing(NR*BUFSIZE, &r1);
d8525fbd	151
422d15ba JA	152	fd = open("largefile", O_RDONLY);
	153	if (fd < 0)
	154	return error("largefile");
d8525fbd	155
422d15ba JA	156	start_timing("Read/write loop");
	157	for (i = 0; i < NR; i++) {
	158	if (read(fd, buffer, BUFSIZE) != BUFSIZE)
	159	return error("largefile read");
	160	write(sk, buffer, BUFSIZE);
d8525fbd	161	}
422d15ba JA	162	end_timing(NR*BUFSIZE, &r2);
	163	close(fd);
	164	close(sk);
d8525fbd	165
422d15ba	166	start_timing("sendfile");
d8525fbd	167	sendfile_again:
422d15ba JA	168	sk = socket(PF_INET, SOCK_STREAM, 0);
	169	if (connect(sk, (struct sockaddr *)&s_to, len) < 0)
	170	return error("connect");
	171
	172	fd = open("largefile", O_RDONLY);
	173	if (fd < 0)
	174	return error("largefile");
	175
	176	i = NR*BUFSIZE;
	177	do {
	178	int ret = sendfile(sk, fd, NULL, i);
	179	i -= ret;
	180	} while (i);
	181
	182	close(fd);
	183	close(sk);
	184	if (--sendfile_loops)
	185	goto sendfile_again;
	186	c1 = end_timing(NRBUFSIZESENDFILE_LOOPS, &r3);
	187
	188	start_timing("splice-pipe");
d8525fbd	189	splice_pipe_again:
422d15ba JA	190	sk = socket(PF_INET, SOCK_STREAM, 0);
	191	if (connect(sk, (struct sockaddr *)&s_to, len) < 0)
	192	return error("connect");
	193
	194	fd = open("largefile", O_RDONLY);
	195	if (fd < 0)
	196	return error("largefile");
	197	if (pipe(pipefd) < 0)
	198	return error("pipe");
	199
	200	i = NR*BUFSIZE;
	201	off = 0;
	202	do {
	203	int ret = splice(fd, &off, pipefd[1], NULL, min(i, BUFSIZE), SPLICE_F_NONBLOCK);
	204	if (ret <= 0)
	205	return error("splice-pipe-in");
	206	i -= ret;
	207	while (ret > 0) {
	208	int flags = i ? SPLICE_F_MORE : 0;
	209	int written = splice(pipefd[0], NULL, sk, NULL, ret, flags);
	210	if (written <= 0)
	211	return error("splice-pipe-out");
	212	ret -= written;
	213	}
	214	} while (i);
	215
	216	close(fd);
	217	close(sk);
	218	close(pipefd[0]);
	219	close(pipefd[1]);
	220	if (--splice_pipe_loops)
	221	goto splice_pipe_again;
	222	c2 = end_timing(NRBUFSIZESPLICE_LOOPS, &r4);
	223
	224	/*
	225	* Direct splicing was disabled as being immediately available,
	226	* it's reserved for sendfile emulation now.
	227	*/
	228	#if 0
	229	start_timing("splice");
d8525fbd	230	splice_again:
422d15ba JA	231	sk = socket(PF_INET, SOCK_STREAM, 0);
	232	if (connect(sk, (struct sockaddr *)&s_to, len) < 0)
	233	return error("connect");
	234
	235	fd = open("largefile", O_RDONLY);
	236	if (fd < 0)
	237	return error("largefile");
	238
	239	i = NR*BUFSIZE;
	240	off = 0;
	241	do {
	242	int flags = BUFSIZE < i ? SPLICE_F_MORE : 0;
	243	int ret;
	244
	245	ret = splice(fd, &off, sk, NULL, min(i, BUFSIZE), flags);
	246
	247	if (ret <= 0)
	248	return error("splice");
	249	i -= ret;
	250	} while (i);
	251
	252	close(fd);
	253	close(sk);
	254	if (--splice_loops)
	255	goto splice_again;
	256	c3 = end_timing(NRBUFSIZESPLICE_LOOPS, &r5);
	257	#else
	258	c3 = 0;
	259	#endif
d8525fbd JA	260
	261	/*
	262	* c1/r3 - sendfile
	263	* c2/r4 - splice-pipe
	264	* c3/r5 - splice
	265	*/
	266
	267	if (c1 && c2)
	268	printf("sendfile is %.2f%% more efficient than splice-pipe.\n",
	269	(c2 - c1) / c1 * 100.0 );
	270	if (c1 && c3)
	271	printf("sendfile is %.2f%% more efficient than splice.\n",
	272	(c3 - c1) / c1 * 100.0 );
	273	if (c2 && c3)
	274	printf("splice is %.2f%% more efficient splice-pipe.\n",
	275	(c2 - c3) / c3 * 100.0 );
	276	if (r3 && r4)
	277	printf("sendfile is %.2f%% faster than splice-pipe.\n",
	278	(r3 - r4) / r4 * 100.0 );
	279	if (r3 && r5)
	280	printf("sendfile is %.2f%% faster than splice.\n",
	281	(r3 - r5) / r5 * 100.0 );
	282	if (r4 && r5)
	283	printf("splice is %.2f%% faster than splice-pipe.\n",
	284	(r5 - r4) / r4 * 100.0 );
	285
	286	return 0;
	287	}
	288
	289
	290	static void setup_shared_var(void)
	291	{
	292	char zerobuff [4096] = { 0, };
	293	int ret, fd;
	294
	295	fd = creat(".tmp_mmap", 0700);
	296	BUG_ON(fd == -1);
	297	close(fd);
	298
	299	fd = open(".tmp_mmap", O_RDWR\|O_CREAT\|O_TRUNC);
	300	BUG_ON(fd == -1);
	301	ret = write(fd, zerobuff, 4096);
	302	BUG_ON(ret != 4096);
	303
	304	cycles = (void *)mmap(0, 4096, PROT_READ\|PROT_WRITE, MAP_SHARED, fd, 0);
	305	BUG_ON(cycles == (void *)-1);
	306
	307	close(fd);
	308	}
	309
	310	#define SCHED_BATCH 3
	311
0d53c23c JA	312	#if defined(__i386__)
0d53c23c JA	313	#define rdtscll(val) \
d8525fbd	314	do { \
0d53c23c	315	__asm__ __volatile__("rdtsc" : "=A" (val)); \
d8525fbd	316	} while (0)
0d53c23c JA	317	#if 0
	318	#elif defined(__ia64__)
	319	#define rdtscll(val) \
d8525fbd	320	do { \
0d53c23c	321	val = *__mm_clock_dev; \
d8525fbd	322	} while (0)
0d53c23c	323	#endif
d8525fbd	324	#else
0d53c23c	325	#define rdtscll(val) \
d8525fbd JA	326	do { (val) = 0LL; } while (0)
	327	#endif
	328
	329	/*
	330	* Keep lowprio looping - to meausure the number of idle cycles
	331	* available. It's tricky: we do a series of RDTSC calls, and
	332	* if the delay to the last measurement was less than 500 cycles,
	333	* we conclude that only this loop ran.
	334	*/
	335	static void lowprio_cycle_soak_loop(void)
	336	{
	337	struct sched_param p = { sched_priority: 0 };
	338	unsigned long long t0, t1, delta;
	339
	340	/*
	341	* We are a nice +19 SCHED_BATCH task:
	342	*/
	343	BUG_ON(sched_setscheduler(0, SCHED_BATCH, &p) != 0);
	344	nice(40);
	345
	346	rdtscll(t0);
	347	while (cycles >= 0) {
	348	rdtscll(t1);
	349	delta = t1-t0;
	350	if (delta < 500)
	351	*cycles += delta;
	352	t0 = t1;
	353	}
	354	}
	355
	356	int main(int argc, char **argv)
	357	{
	358	unsigned int sk, len;
	359	struct sockaddr addr;
	360	pid_t pid;
	361
	362	setup_shared_var();
	363
	364	signal(SIGCHLD, SIG_IGN);
	365	sk = socket(PF_INET, SOCK_STREAM, 0);
	366	if (sk < 0) {
	367	perror("socket");
	368	exit(1);
	369	}
	370	if (listen(sk, 1) < 0) {
	371	perror("listen");
	372	exit(1);
	373	}
	374	len = sizeof(addr);
	375	if (getsockname(sk, &addr, &len) < 0) {
	376	perror("getsockname");
	377	exit(1);
	378	}
	379	pid = fork();
	380	if (!pid) {
	381	lowprio_cycle_soak_loop();
	382	exit(0);
	383	}
	384	nice(-20);
	385	child(&addr, len);
	386	kill(pid, SIGHUP);
	387	exit(0);
	388	}