From owner-freebsd-current  Mon Jul 12 10:41:57 1999
Delivered-To: freebsd-current@freebsd.org
Received: from apollo.backplane.com (apollo.backplane.com [209.157.86.2])
	by hub.freebsd.org (Postfix) with ESMTP id 8568315135
	for <freebsd-current@FreeBSD.ORG>; Mon, 12 Jul 1999 10:41:49 -0700 (PDT)
	(envelope-from dillon@apollo.backplane.com)
Received: (from dillon@localhost)
	by apollo.backplane.com (8.9.3/8.9.1) id KAA70496;
	Mon, 12 Jul 1999 10:41:10 -0700 (PDT)
	(envelope-from dillon)
Date: Mon, 12 Jul 1999 10:41:10 -0700 (PDT)
From: Matthew Dillon <dillon@apollo.backplane.com>
Message-Id: <199907121741.KAA70496@apollo.backplane.com>
To: Matthew Dillon <dillon@apollo.backplane.com>
Cc: Poul-Henning Kamp <phk@critter.freebsd.dk>,
	Luoqi Chen <luoqi@watermarkgroup.com>, dfr@nlsystems.com,
	jeremyp@gsmx07.alcatel.com.au, freebsd-current@FreeBSD.ORG,
	mike@ducky.net
Subject: lock prefix overhead in SMP system - results
Sender: owner-freebsd-current@FreeBSD.ORG
Precedence: bulk
X-Loop: FreeBSD.ORG

    Ok, here we are.  The lock instruction overhead is significant relative
    the execution time of the instruction, though on a 450 MHz P-III it is
    still not a huge problem.  Basically 12 vs 89 ns for the non competitive
    case, and 26 vs 161 ns for the competitive case.

empty loop	  9.33 ns/loop	mode 0
1proc no locks	 12.86 ns/loop	mode 9
2proc no locks   26.16 ns/loop	mode 10
1proc w/ locks	 89.87 ns/loop  mode 1
2proc w/ locks  161.09 ns/loop	mode 2

    If you modify the source code and add three nop() calls inside the loop
    for case 9 and 1 (which also covers 10 and 2), which avoids some of the
    synchronization overhead, you get numbers closer to what I expect in the
    real world.  ( recompile the code -DREAL_WORLD )

empty loop	  9.33 ns/loop	mode 0
1proc no locks	 55.89 ns/loop
2proc no locks  129.96 ns/loop
1proc w/ locks   98.35 ns/loop
2proc w/ locks  160.96 ns/loop


/*
 * Compile -O2
 */

#include <sys/types.h>
#include <sys/mman.h>
#include <sys/time.h>
#include <stdio.h>
#include <stdlib.h>
#include <fcntl.h>
#include <stdarg.h>
#include <unistd.h>

#define LOOPS	10000000

#define ATOMIC_ASM(type,op)     \
    __asm __volatile ("lock; " op : "=m" (*(type *)p) : "ir" (v), "0" (*(type *)p))

#define ATOMIC_ASM_NOLOCK(type,op)     \
    __asm __volatile (op : "=m" (*(type *)p) : "ir" (v), "0" (*(type *)p))

static __inline void
atomic_add_int(void *p, u_int v)
{
        ATOMIC_ASM(int, "addl %1,%0");
}

static __inline void
atomic_add_int_nolock(void *p, u_int v)
{
        ATOMIC_ASM_NOLOCK(int, "addl %1,%0");
}

void
nop(void)
{
}

int
main(int ac, char **av)
{
    int fd;
    char *ptr;
    char *wlocks;
    int pgsize = getpagesize();
    volatile int i;
    int m;
    int usec;
    pid_t pid = -1;
    struct timeval tv1;
    struct timeval tv2;

    if (ac == 1) {
	printf("%s {0,1,2}\n", av[0]);
	printf(" 0 - empty loop\n");
	printf(" 1 - one process\n");
	printf(" 2 - two processes\n");
	printf(" 8 - empty loop - no lock\n");
	printf(" 9 - one process - no lock\n");
	printf(" 10- two processes - no lock\n");
	exit(1);
    }
    m = strtol(av[1], NULL, 0);

    fd = open("test.dat", O_RDWR|O_CREAT|O_TRUNC, 0666);
    ftruncate(fd, pgsize);
    ptr = mmap(NULL, pgsize, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);

    if (m & 8)
	wlocks = "";
    else
	wlocks = " w/locks";

    gettimeofday(&tv1, NULL);
    switch(m) {
    case 10:
	pid = fork();
	/* fall through */
    case 9:
	for (i = 0; i < LOOPS; ++i) {
	    atomic_add_int_nolock(ptr, 1);
#ifdef REAL_WORLD
	    nop();
	    nop();
	    nop();
#endif
	}
	break;
    case 8:
	for (i = 0; i < LOOPS; ++i) {
		;
	}
	break;
    case 2:
	pid = fork();
	/* fall through */
    case 1:
	for (i = 0; i < LOOPS; ++i) {
	    atomic_add_int(ptr, 1);
#ifdef REAL_WORLD
	    nop();
	    nop();
	    nop();
#endif
	}
	break;
    case 0:
	for (i = 0; i < LOOPS; ++i) {
		;
	}
	break;
    default:
	printf("huh?\n");
	exit(1);
    }
    if (pid == 0)
	_exit(0);
    while (wait(NULL) > 0)
	;
    gettimeofday(&tv2, NULL);

    usec = tv2.tv_usec + 1000000 - tv1.tv_usec + (tv2.tv_sec - tv1.tv_sec - 1) * 1000000;

    switch(m) {
    case 0:
    case 8:
	printf("empty%s %4.2f ns/loop\n", wlocks, (double)usec * 1000.0 / (double)LOOPS);
	break;
    case 1:
    case 9:
	printf("1proc%s %4.2f ns/loop\n", wlocks, (double)usec * 1000.0 / (double)LOOPS);
	break;
    case 2:
    case 10:
	printf("2proc%s %4.2f ns/loop\n", wlocks, (double)usec * 1000.0 / (double)LOOPS / 2.0);
	break;
    }
    return(0);
}



To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-current" in the body of the message