Skip to contentSkip to conversation

latency error (~2ms) with nanosleep

DISCUSSION quade 5 participants 6 messages Jun 13, 2005 – Jun 14, 2005

lore
Share & tools
lore

quade

author
#0 lore raw

Playing around with the (simple) measurement of latency-times I noticed, that the systemcall "nanosleep" has always a minimal latency from about ~2ms (haven't run it all night, so...). It seems to be a systematical error.

A short investigation shows, that "sys_nanosleep()" uses schedule_timeout(), but schedule_timeout() is working exactly as expected. Therefore I think it has something to do with the scheduling?

Has someone an explanation for the ~2ms error?
If it is indeed a systematical error, does it make sense to "adjust" (correct) this error in the systemcall "sys_nanosleep()"?

Find attached my small test program.

Juergen.

#include <stdio.h>
#include <time.h>
#include <sched.h>
#include <sys/time.h>

#define TAKTFREQ 600

static inline unsigned long long int rdtsc()

{
	unsigned long long int x;
	__asm__ volatile (".byte 0x0f, 0x31" : "=A" (x));
	return x;
}

// XXX - I know, there can be an overrun ...
static inline unsigned long time_in_usec( struct timeval *tv )

{
	return (tv->tv_sec*1000000)+tv->tv_usec;
}

int main( int argc, char **argv )

{
	int i;
	struct timespec delay;
	struct timeval tvstart, tvend;
	unsigned long start, end, timediff, maxdiff, shouldbetime;
	unsigned long mindiff;
#if 1
	struct sched_param SchedulingParameter;

	SchedulingParameter.sched_priority = 50;
	if( sched_setscheduler( 0, SCHED_RR, &SchedulingParameter )!= 0 ) {
		perror( "Set Scheduling Priority" );
		return -1;
	}

#endif

	for( shouldbetime=1000; shouldbetime<51000;shouldbetime+=1000 ) {
		mindiff = 0xffffffff;
		maxdiff = 0;
		for( i=0; i<30; i++ ) {
			delay.tv_sec  = 0;
			delay.tv_nsec = shouldbetime*1000; // in nsec
			gettimeofday(&tvstart,NULL);
			start = rdtsc();
			nanosleep( &delay, NULL );
			end = rdtsc();
			gettimeofday(&tvend,NULL);
#if 0
			printf("timdiff: %ld - %ld\n",
				(end-start)/TAKTFREQ,  // 600 MHz - in usec
				time_in_usec(&tvend)-time_in_usec(&tvstart));
#endif
			timediff = time_in_usec(&tvend)-time_in_usec(&tvstart);
			timediff -= shouldbetime;
			if( (timediff > maxdiff)&&i>0 )
				maxdiff = timediff;
			if( (timediff < mindiff)&&i>0 )
				mindiff = timediff;
		}
		//printf("%7.1ld-> max diff: %ld\n", shouldbetime, maxdiff );
		printf("%7.1ld %ld %ld\n", shouldbetime, maxdiff, mindiff );
	}
	return 0;
}
#1 lore raw

nanosleep, according to the documation is supposed to sleep "at least" the 'struct timespec' time. It can return in a shorter time as a result of a signal and, if so, the input time-values will be updated accordingly. The resolution is limited to the HZ value. This means that it will, unless interrupted, always sleep at least 1 / HZ seconds (about 1 ms on current x86 distributions).

FYI, there is no 'fine resolution' timer available on any Linux-ported platform that could take advantage of the nanosecond input resolution of the function.

Quoted 11 lines · quade

Playing around with the (simple) measurement of latency-times I noticed, that the systemcall "nanosleep" has always a minimal latency from about ~2ms (haven't run it all night, so...). It seems to be a systematical error.

A short investigation shows, that "sys_nanosleep()" uses schedule_timeout(), but schedule_timeout() is working exactly as expected. Therefore I think it has something to do with the scheduling?

Quoted 8 lines

Has someone an explanation for the ~2ms error?
If it is indeed a systematical error, does it make sense to "adjust" (correct) this error in the systemcall "sys_nanosleep()"?

Find attached my small test program.

Juergen.

Cheers,
Dick Johnson
Penguin : Linux version 2.6.11.9 on an i686 machine (5537.79 BogoMips).
Notice : All mail here is now cached for review by Dictator Bush.
98.36% of all statistics are fiction.

#2 lore raw
Quoted 9 lines · Richard B. Johnson wrote/

nanosleep, according to the documation is supposed to sleep "at least" the 'struct timespec' time. It can return in a shorter time as a result of a signal and, if so, the input time-values will be updated accordingly. The resolution is limited to the HZ value. This means that it will, unless interrupted, always sleep at least 1 / HZ seconds (about 1 ms on current x86 distributions).

Quoted 4 lines

FYI, there is no 'fine resolution' timer available on any Linux-ported platform that could take advantage of the nanosecond input resolution of the function.

I could add that there is a project working on this issue: high-resolution timers . You can download patches for 2.6 and 2.4 . The precision is in the order of 10 or 100 µs.

Eric

#3 lore raw
In reply to quade
Quoted 4 lines · quade

Playing around with the (simple) measurement of latency-times I noticed, that the systemcall "nanosleep" has always a minimal latency from about ~2ms (haven't run it all night, so...). It seems to be a systematical error.

Known issue. The x86 interrupt usually has a period of slightly less than a ms. It will therefore generally add nearly a whole ms to ensure that it does not ever wait for *less* than specified.

Chris

#4 lore raw
Quoted 9 lines · Chris Friesen
Quoted 4 lines · quade

Playing around with the (simple) measurement of latency-times I noticed, that the systemcall "nanosleep" has always a minimal latency from about ~2ms (haven't run it all night, so...). It seems to be a systematical error.

Known issue. The x86 interrupt usually has a period of slightly less than a ms. It will therefore generally add nearly a whole ms to ensure that it does not ever wait for *less* than specified.

Exactly. And the sys_nanosleep() code adds one more if the parameter has any positive value at all:

        expire = timespec_to_jiffies(&t) + (t.tv_sec || t.tv_nsec);
        current->state = TASK_INTERRUPTIBLE;
        expire = schedule_timeout(expire);

Thanks,
Nish

author
#5 lore raw
Quoted 27 lines · Nish Aravamudan
Quoted 9 lines · Chris Friesen
Quoted 4 lines · quade

Playing around with the (simple) measurement of latency-times I noticed, that the systemcall "nanosleep" has always a minimal latency from about ~2ms (haven't run it all night, so...). It seems to be a systematical error.

Known issue. The x86 interrupt usually has a period of slightly less than a ms. It will therefore generally add nearly a whole ms to ensure that it does not ever wait for *less* than specified.

Exactly. And the sys_nanosleep() code adds one more if the parameter has any positive value at all:

        expire = timespec_to_jiffies(&t) + (t.tv_sec || t.tv_nsec);


        current->state = TASK_INTERRUPTIBLE;
        expire = schedule_timeout(expire);

Thanks,
Nish
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/