[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*1024U)

#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(void)
{
	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;
	unsigned int i;
	int 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, sizeof(s_to)) < 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, sizeof(s_to)) < 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, sizeof(s_to)) < 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, sizeof(s_to)) < 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(__attribute__((__unused__)) int argc, __attribute__((__unused__)) char **argv)
{
	pid_t pid;

	setup_shared_var();

	signal(SIGCHLD, SIG_IGN);

	pid = fork();
	if (!pid) {
		lowprio_cycle_soak_loop();
		exit(0);
	}

	nice(-20);
	child();
	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)
000b72a8	28	#define BUFSIZE (64*1024U)
d8525fbd JA	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
000b72a8	115	static int child(void)
d8525fbd JA	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;
000b72a8 JA	124	unsigned int i;
000b72a8 JA	125	int pipefd[2];
d8525fbd JA	126	loff_t off = 0;
	127
	128	r1 = r2 = r3 = r4 = r5 = 0;
	129
	130	sk = socket(PF_INET, SOCK_STREAM, 0);
	131	if (!sk)
	132	return error("socket");
	133	hp = gethostbyname (TARGET_HOSTNAME);
	134	BUG_ON(!hp);
	135	bzero ((char *) &s_to, sizeof (s_to));
	136	bcopy ((char ) hp->h_addr, (char ) &(s_to.sin_addr), hp->h_length);
	137	s_to.sin_family = hp->h_addrtype;
	138	s_to.sin_port = htons(1111);
	139
	140	calibrate_loops();
	141
	142	fprintf(stdout, "BUFSIZE = %d\n", BUFSIZE);
	143	fflush(stdout);
	144
000b72a8	145	if (connect(sk, (struct sockaddr *)&s_to, sizeof(s_to)) < 0)
422d15ba	146	return error("connect");
d8525fbd	147
422d15ba JA	148	start_timing("Empty buffer");
	149	for (i = 0; i < NR; i++)
	150	write(sk, buffer, BUFSIZE);
	151	end_timing(NR*BUFSIZE, &r1);
d8525fbd	152
422d15ba JA	153	fd = open("largefile", O_RDONLY);
	154	if (fd < 0)
	155	return error("largefile");
d8525fbd	156
422d15ba JA	157	start_timing("Read/write loop");
	158	for (i = 0; i < NR; i++) {
	159	if (read(fd, buffer, BUFSIZE) != BUFSIZE)
	160	return error("largefile read");
	161	write(sk, buffer, BUFSIZE);
d8525fbd	162	}
422d15ba JA	163	end_timing(NR*BUFSIZE, &r2);
	164	close(fd);
	165	close(sk);
d8525fbd	166
422d15ba	167	start_timing("sendfile");
d8525fbd	168	sendfile_again:
422d15ba	169	sk = socket(PF_INET, SOCK_STREAM, 0);
000b72a8	170	if (connect(sk, (struct sockaddr *)&s_to, sizeof(s_to)) < 0)
422d15ba JA	171	return error("connect");
	172
	173	fd = open("largefile", O_RDONLY);
	174	if (fd < 0)
	175	return error("largefile");
	176
	177	i = NR*BUFSIZE;
	178	do {
	179	int ret = sendfile(sk, fd, NULL, i);
	180	i -= ret;
	181	} while (i);
	182
	183	close(fd);
	184	close(sk);
	185	if (--sendfile_loops)
	186	goto sendfile_again;
	187	c1 = end_timing(NRBUFSIZESENDFILE_LOOPS, &r3);
	188
	189	start_timing("splice-pipe");
d8525fbd	190	splice_pipe_again:
422d15ba	191	sk = socket(PF_INET, SOCK_STREAM, 0);
000b72a8	192	if (connect(sk, (struct sockaddr *)&s_to, sizeof(s_to)) < 0)
422d15ba JA	193	return error("connect");
	194
	195	fd = open("largefile", O_RDONLY);
	196	if (fd < 0)
	197	return error("largefile");
	198	if (pipe(pipefd) < 0)
	199	return error("pipe");
	200
	201	i = NR*BUFSIZE;
	202	off = 0;
	203	do {
	204	int ret = splice(fd, &off, pipefd[1], NULL, min(i, BUFSIZE), SPLICE_F_NONBLOCK);
	205	if (ret <= 0)
	206	return error("splice-pipe-in");
	207	i -= ret;
	208	while (ret > 0) {
	209	int flags = i ? SPLICE_F_MORE : 0;
	210	int written = splice(pipefd[0], NULL, sk, NULL, ret, flags);
	211	if (written <= 0)
	212	return error("splice-pipe-out");
	213	ret -= written;
	214	}
	215	} while (i);
	216
	217	close(fd);
	218	close(sk);
	219	close(pipefd[0]);
	220	close(pipefd[1]);
	221	if (--splice_pipe_loops)
	222	goto splice_pipe_again;
	223	c2 = end_timing(NRBUFSIZESPLICE_LOOPS, &r4);
	224
	225	/*
	226	* Direct splicing was disabled as being immediately available,
	227	* it's reserved for sendfile emulation now.
	228	*/
	229	#if 0
	230	start_timing("splice");
d8525fbd	231	splice_again:
422d15ba	232	sk = socket(PF_INET, SOCK_STREAM, 0);
000b72a8	233	if (connect(sk, (struct sockaddr *)&s_to, sizeof(s_to)) < 0)
422d15ba JA	234	return error("connect");
	235
	236	fd = open("largefile", O_RDONLY);
	237	if (fd < 0)
	238	return error("largefile");
	239
	240	i = NR*BUFSIZE;
	241	off = 0;
	242	do {
	243	int flags = BUFSIZE < i ? SPLICE_F_MORE : 0;
	244	int ret;
	245
	246	ret = splice(fd, &off, sk, NULL, min(i, BUFSIZE), flags);
	247
	248	if (ret <= 0)
	249	return error("splice");
	250	i -= ret;
	251	} while (i);
	252
	253	close(fd);
	254	close(sk);
	255	if (--splice_loops)
	256	goto splice_again;
	257	c3 = end_timing(NRBUFSIZESPLICE_LOOPS, &r5);
	258	#else
	259	c3 = 0;
	260	#endif
d8525fbd JA	261
	262	/*
	263	* c1/r3 - sendfile
	264	* c2/r4 - splice-pipe
	265	* c3/r5 - splice
	266	*/
	267
	268	if (c1 && c2)
	269	printf("sendfile is %.2f%% more efficient than splice-pipe.\n",
	270	(c2 - c1) / c1 * 100.0 );
	271	if (c1 && c3)
	272	printf("sendfile is %.2f%% more efficient than splice.\n",
	273	(c3 - c1) / c1 * 100.0 );
	274	if (c2 && c3)
	275	printf("splice is %.2f%% more efficient splice-pipe.\n",
	276	(c2 - c3) / c3 * 100.0 );
	277	if (r3 && r4)
	278	printf("sendfile is %.2f%% faster than splice-pipe.\n",
	279	(r3 - r4) / r4 * 100.0 );
	280	if (r3 && r5)
	281	printf("sendfile is %.2f%% faster than splice.\n",
	282	(r3 - r5) / r5 * 100.0 );
	283	if (r4 && r5)
	284	printf("splice is %.2f%% faster than splice-pipe.\n",
	285	(r5 - r4) / r4 * 100.0 );
	286
	287	return 0;
	288	}
	289
	290
	291	static void setup_shared_var(void)
	292	{
	293	char zerobuff [4096] = { 0, };
	294	int ret, fd;
	295
	296	fd = creat(".tmp_mmap", 0700);
	297	BUG_ON(fd == -1);
	298	close(fd);
	299
	300	fd = open(".tmp_mmap", O_RDWR\|O_CREAT\|O_TRUNC);
	301	BUG_ON(fd == -1);
	302	ret = write(fd, zerobuff, 4096);
	303	BUG_ON(ret != 4096);
	304
	305	cycles = (void *)mmap(0, 4096, PROT_READ\|PROT_WRITE, MAP_SHARED, fd, 0);
	306	BUG_ON(cycles == (void *)-1);
	307
	308	close(fd);
	309	}
	310
	311	#define SCHED_BATCH 3
	312
0d53c23c JA	313	#if defined(__i386__)
0d53c23c JA	314	#define rdtscll(val) \
d8525fbd	315	do { \
0d53c23c	316	__asm__ __volatile__("rdtsc" : "=A" (val)); \
d8525fbd	317	} while (0)
0d53c23c JA	318	#if 0
	319	#elif defined(__ia64__)
	320	#define rdtscll(val) \
d8525fbd	321	do { \
0d53c23c	322	val = *__mm_clock_dev; \
d8525fbd	323	} while (0)
0d53c23c	324	#endif
d8525fbd	325	#else
0d53c23c	326	#define rdtscll(val) \
d8525fbd JA	327	do { (val) = 0LL; } while (0)
	328	#endif
	329
	330	/*
	331	* Keep lowprio looping - to meausure the number of idle cycles
	332	* available. It's tricky: we do a series of RDTSC calls, and
	333	* if the delay to the last measurement was less than 500 cycles,
	334	* we conclude that only this loop ran.
	335	*/
	336	static void lowprio_cycle_soak_loop(void)
	337	{
	338	struct sched_param p = { sched_priority: 0 };
	339	unsigned long long t0, t1, delta;
	340
	341	/*
	342	* We are a nice +19 SCHED_BATCH task:
	343	*/
	344	BUG_ON(sched_setscheduler(0, SCHED_BATCH, &p) != 0);
	345	nice(40);
	346
	347	rdtscll(t0);
	348	while (cycles >= 0) {
	349	rdtscll(t1);
	350	delta = t1-t0;
	351	if (delta < 500)
	352	*cycles += delta;
	353	t0 = t1;
	354	}
	355	}
	356
000b72a8	357	int main(__attribute__((__unused__)) int argc, __attribute__((__unused__)) char **argv)
d8525fbd	358	{
d8525fbd JA	359	pid_t pid;
	360
	361	setup_shared_var();
	362
	363	signal(SIGCHLD, SIG_IGN);
000b72a8	364
d8525fbd JA	365	pid = fork();
	366	if (!pid) {
	367	lowprio_cycle_soak_loop();
	368	exit(0);
	369	}
000b72a8	370
d8525fbd	371	nice(-20);
000b72a8	372	child();
d8525fbd JA	373	kill(pid, SIGHUP);
	374	exit(0);
	375	}