[LITMUS^RT] Interesting thread cpu time VS wall clock

[Glenn Elliott]

> On Mar 16, 2016, at 10:40 AM, Shuai Zhao <zs673 at york.ac.uk> wrote:
> 
> Hi
> 
> I found an interesting thing while I trying to get response time of tasks: The CPU time of a task (CLOCK_THREAD_CPUTIME_ID) accounts faster than wall clock (CLOCK_REALTIME) ! 
> 
> I noticed that you use the same thing in the rtspin.c so I guess that you may know the answer.
> 
> I know that this is a question related to generic Linux or may be the cpu clock speed instead of litmus, but would you please have a look and help me?
> 
> The CPU I use is i7 6700k 4.0MHz. But the clock speed is actually 4006 MHz when I cat /proc/cpuinfo. 
> 
> Here is the code:
> 
> /************* CPU cycle cosuming for 3 threads *************/
> #define NUMS 500
> static int num[NUMS];
> static int loop_once(void) {
> 	int i, j = 0;
> 	for (i = 0; i < NUMS; i++)
> 		j += num[i]++;
> 	return j;
> }
> /************* CPU cycle cosuming END *************/
> 
> void* run(void *tcontext) {
> 	struct timespec start, end, start1, end1, sleep = {0,10};
> 	int i;
> 
> 	exit_program = 0;
> 
> 	do {
> 		nanosleep(&sleep, NULL);
> 
> 		clock_gettime(CLOCK_REALTIME, &start);
> 		clock_gettime(CLOCK_THREAD_CPUTIME_ID, &start1);
> 
> 
> 		for(i = 0; i < loop_times; i++)
> 			loop_once();
> 
> 
> 		clock_gettime(CLOCK_THREAD_CPUTIME_ID, &end1);
> 		clock_gettime(CLOCK_REALTIME, &end);	
> 
> 									
> 		response_time[0][count_first] = (end.tv_sec * 1000000000 + end.tv_nsec) - (start.tv_sec * 1000000000 + start.tv_nsec);
> 
> 		response_time[1][count_first] = (end1.tv_sec * 1000000000 + end1.tv_nsec) - (start1.tv_sec * 1000000000 + start1.tv_nsec);
> 
> 
> 		/*finish the program */
> 		count_first++;
> 		if (count_first >= 10000)
> 			exit_program = 1;
> 	} while (exit_program != 1);
> 
> 	return NULL;
> }
> 
> I basically copies the code from rtspin.c here. In the program I get time stamps of the wall-clock and thread-cpu clock and then let the thread loops for a given times. Then I get time stamps again to check the differences between these two clocks (wallclock - threadcpuclock).
> 
> The difference I expect should be the interval of wall-clock always be bigger than thread cpu clock as there will be timer or reschedule interrupts.
> 
> But on my machine I got amazing results: with the increment of the loop_times, the thread-cpu clock will goes faster than wall clock and I got results like this: (clock 0 is wall clock and clock 1 is thread cpu clock)
> 
> <IMG_0257.JPG>​
> 
> So we can see that the thread cpu clock goes faster than wall clock by 33191 nanosecond when we loop 100000 times and I am shock by this result.
> 
> If the wall clock is greater then it makes sense, but here we have a negative difference. 
> 
> One possibility is that the wall clock and the cpu time are accounted by different timers in kernel. But they should all updated based on the HZ, which is 1000 in my case. So I cannot explain the reason why the cpu time in my computer counts faster than the wall clock.
> 
> Thank you in advance for helping and looking forward your reply.
> 
> Best wishes
> Shuai

Hi Shuai,

Could NTP or PTP be responsible?  Seems unlikely, but can you rule it out?

-Glenn