Interesting question regarding the process in Linux

namanhams · 03-23-2008, 07:31 AM

Hi all,

I have a question about process in Linux.
First of all, i have a C program, that run fork() every 1 seconds, and the child processes also do that recursively.

Here's the code :

Code:

#include <stdio.h>
#include <unistd.h>

int main()

{

	setlinebuf(stdout);

	printf("ORIGINAL PID = %d\n", getpid());

	while (1) {

		if (!fork()) printf("NEW PID = %d\n", getpid());

		sleep(1); // wait 1s, so forking is slowed down

	}

}

If i just run the program, the system hangs. It may be because of too many process running. So i set the limit number of processes to , says 1000, by :
ulimit -u 1000

Then i run the program again. Ok, now it works.

The problems is, when i set the time to be smaller than 1 second, then the system runs very very slowly. Even when i set the time to 0.99 second, it's still the same.

The question is : how come does it happen ? Why the difference between 1 and 0.99 can make such a big change ?

Hope to hear some explain here. Thanks.

Hko · 03-23-2008, 07:40 AM

I am not sure, but what I think is this:

With one second timer, the system/kernel checks each of your 1000 processes once a second. But when you specify 0.99 second, it uses high resolution timers and check all the processes each 0.01 secs, and the system load gets 100 times larger.

I agree it is interesting...

ophirg · 03-23-2008, 07:55 AM

I think your program doesn't start a new process every second.
The number of processes over time is not linear, it's exponential.
Each second, every process starts it's own program. So in second N you'll have about 2^N processes.
If you want linear number of processes you should use:
...
if (fork()==0) {
printf("NEW PID = %d\n", getpid());
return 0;
}
else sleep(1);
...

embesil · 03-23-2008, 01:12 PM

....edit....

namanhams · 03-24-2008, 01:02 AM

@ Hko : Can you explain in more detailed ? So if i set the time to 0.1 second, then the system check all the processes each 0.9 second or 0.01 second ?

@ophirg : Yes, you'r correct. Thanks.

Any more explaination ?

Hko · 03-24-2008, 05:42 AM

Quote:

Originally Posted by namanhams

@ Hko : Can you explain in more detailed ? So if i set the time to 0.1 second, then the system check all the processes each 0.9 second or 0.01 second ?

Any more explaination ?

No, that is not what I meant. I thought the kernel might check your processes every 1 sec when you pause/sleep for 1 sec. But when you pause for 0.99 it starts to use high-precision timing. Like when you pause for 0.9 maybe it checks every 0.1 second or so.

But now I had a look at your program a second time, I think it is different. How did you sleep for 0.99 ? With sleep(0.99)? If so, then there is a much simpler explanation: The sleep() function only takes an integer, so your 0.99 maybe rounded down to zero...

namanhams · 03-24-2008, 07:59 AM

Quote:

Originally Posted by Hko

But now I had a look at your program a second time, I think it is different. How did you sleep for 0.99 ? With sleep(0.99)? If so, then there is a much simpler explanation: The sleep() function only takes an integer, so your 0.99 maybe rounded down to zero...

I look at the "man sleep" page but it doesn't say that the time should be integer. How do you know ?

Hko · 03-24-2008, 10:07 AM

Quote:

Originally Posted by namanhams

I look at the "man sleep" page but it doesn't say that the time should be integer. How do you know ?

Actually the man page really does say the time should be integer:

Quote:

Code:

SYNOPSIS
       #include <unistd.h>

       unsigned int sleep(unsigned int seconds);

If yours did not, you were probably looking at the man page of the sleep shell command. Try "man 3 sleep" to be sure you the man page about the C-library function sleep().