[[email protected]: process creation time increases linearly with shmem]

----- Forwarded message from Ray Fucillo <[email protected]> -----

Subject: process creation time increases linearly with shmem
From: Ray Fucillo <[email protected]>
To: [email protected]
Date: Wed, 24 Aug 2005 14:43:29 -0400
Resent-Message-Id: <[email protected]>
Resent-Sender: Benjamin LaHaise <[email protected]>
Resent-From: [email protected]
Resent-Date: Wed, 24 Aug 2005 15:13:51 -0400
Resent-To: [email protected]

I am seeing process creation time increase linearly with the size of the shared memory segment that the parent touches. The attached forktest.c is a very simple user program that illustrates this behavior, which I have tested on various kernel versions from 2.4 through 2.6. Is this a known issue, and is it solvable?

TIA,
Ray

#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <unistd.h>
#include <sys/time.h>
#include <errno.h>
#include <sys/shm.h>
#include <sys/wait.h>
#include <sys/types.h>
#include <signal.h>

#define MAXJOBS 50
#define MAXMALLOC 1024

#define USESIGCHLDHND
/* USESIGCHLDHND feature code changes how the parent waits
for the children. When this feature code is on we define a signal handler for SIGCHLD and call waitpid to clean up the child process. If this feature code is off, we wait until all children are forked and then loop through the array of child pids and call waitpid() on each. The purpose of this feature code was to see if there is any difference in timing based on cleaning up zombies faster. Test have shown no appreciable difference. */

/* Return a floating point number of seconds since the start
time in the timeval structure pointed to by starttv */

float elapsedtime(struct timeval *starttv) {
	struct timeval currenttime;
	gettimeofday(&currenttime,NULL);
	return ((currenttime.tv_sec - starttv->tv_sec) +
	       ((float)(currenttime.tv_usec - starttv->tv_usec)/1000000));
}

#ifdef USESIGCHLDHND
int childexitcnt = 0;

void sigchldhnd(int signum) {
	if (waitpid(-1,NULL,WNOHANG)) ++childexitcnt;
	return;
}

#endif

int main(void) {
	pid_t childpid[MAXJOBS];
	int x,i;
	int childcnt = 0;
        float endfork, endwait;
	struct shmid_ds myshmid_ds;
	unsigned int mb;
	int myshmid;
	key_t mykey = 0xf00df00d;
	char *mymem = 0;
	struct timeval starttime;
#ifdef USESIGCHLDHND
	struct sigaction sa;
	sa.sa_handler = sigchldhnd;
	sigemptyset(&sa.sa_mask);
	sa.sa_flags = SA_RESTART;
	if (sigaction(SIGCHLD, &sa, NULL) == -1) {
	   printf("sigaction() failed, errno %d - exiting\n",errno);
	   exit(1);
	}
#endif
        printf("\nNumber of jobs to fork (max %d):  ",MAXJOBS);
        scanf("%d",&x);
        if ((x < 1) || (x > MAXJOBS)) {
           printf("\ninvalid input - exiting\n");
           exit(1);
        }
        printf("\nNumber of MB to allocate (0-%d):  ",MAXMALLOC);
        scanf("%d",&mb);
        if (mb > MAXMALLOC) {
           printf("\ninvalid input - exiting\n");
           exit(1);
        }
	/* allocate and initialize shared memory if number
	   of MB is not zero */
	if (mb) {
	   myshmid = shmget(mykey,mb*1024*1024,IPC_CREAT|0777);
	   if (myshmid == -1) {
	      printf("\nshmget() failed, errno %d. - exiting\n",errno);
	      exit(1);
	   }
	   mymem = (char *) shmat(myshmid,0,0);
	   if (mymem == (char *) -1) {
	      printf("\nshmat() failed, errno %d. - exiting\n",errno);
	      exit(1);
	   }
	   if (shmctl(myshmid,IPC_STAT,&myshmid_ds)) {
	      printf("\nshmctl() failed, errno %d. - exiting\n",errno);
	      exit(1);
	   }
	   /* write a pattern in the new shmem segment*/
	   for (i=0; i < (mb*1024*1024); i+=32) mymem[i]='R';
	}	
	printf("\nStarting %d jobs.  time:0.0", x);
	fflush(stdout);
	gettimeofday(&starttime,NULL); 
	for (i=0; i<x; i++) {
	   childpid[i] = fork();
	   if (!childpid[i]) {
	      /* child process */
	      printf("\n - Child %d         time:%f",i,elapsedtime(&starttime));
	      exit(1);
	   } else if (childpid[i] == -1) {
	      /* failure */
	      printf("\nfork failed, errno = %d");
	   } else childcnt++;
	}
	endfork = elapsedtime(&starttime);
#ifndef USESIGCHLDHND
	for (i=0; i<x; i++) waitpid(childpid[i],0,0);
#else
	while (childexitcnt < childcnt) {
	   if (waitpid(-1,NULL,0)) ++childexitcnt;
	}
#endif	   
	endwait = elapsedtime(&starttime);
	printf("\nTime to fork all processes in seconds: %f", endfork);
        printf("\nTime for all processes to complete: %f\n", endwait);

	/* kill shmem segment */
	if ((mb) && (shmctl(myshmid,IPC_RMID,&myshmid_ds))) {
	   printf("\nshmctl() failed, errno %d. - exiting\n",errno);
	   exit(1);
	}
}

----- End forwarded message -----
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

FYI, see my reply on lkml too (it might start a thread of its own).

I recently had a quick look into why this is so when I heard some other operating systems do better than us here.

I believe it is because we copy all page tables though we need not copy those that are filebacked MAP_SHARED. The patch to change this is fairly trivial, though I didn't have a real world test to justify any improvement.

Actually, it may be worthwhile for another reason on large NUMA systems so their leaf page tables get allocated on the same node in which the memory is used.

[snip]

SUSE Labs, Novell Inc.

Send instant messages to your online friends http://au.messenger.yahoo.com --
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>