On Thursday 22 June 2006 14:23, Chuck Ebbert wrote:
> In-Reply-To: <200606211914.37137.ak@suse.de>
> 
> On Wed, 21 Jun 2006 19:14:37 +0200, Andi Kleen wrote:
> 
> >> 
> >> /* test how fast lsl/jnz/and runs.
> >>  */
> >> #define _GNU_SOURCE
> >> #include <stdio.h>
> >> #include <stdlib.h>
> >> 
> >> #define rdtscll(t)   asm volatile ("rdtsc" : "=A" (t))
> >> 
> >> #ifndef ITERS
> >> #define ITERS        1000000
> >> #endif
> >> 
> >> int main(int argc, char * const argv[])
> >> {
> >>      unsigned long long tsc1, tsc2;
> >>      int count, cpu, junk;
> >> 
> >>      rdtscll(tsc1);
> >>      asm (
> >>              "       pushl %%ds              \n"
> >>              "       popl %2                 \n"
> >>              "1:                             \n"
> >> #ifdef DO_TEST
> >>              "       lsl %2,%0               \n"
> >>              "       jnz 2f                  \n"
> >>              "       and $0xff,%0            \n"
> >> #endif
> >>              "       dec %1                  \n"
> >>              "       jnz 1b                  \n"
> >>              "2:                             \n"
> >>              : "=&r" (cpu), "=&r" (count), "=&r" (junk)
> >>              : "1" (ITERS), "0" (-1)
> >>      );
> >>      rdtscll(tsc2);
> >
> > Measuring this way is a bad idea because you get far too much 
> > noise from the RDTSCs. Usually you need to put a a few thousands entry 
> > loop inside the RDTSCP and devide the result by the loop count
> 
> I got tired of people (namely me) forgetting to compile the C code
> with optimization, so I did the loop in assembler.  It does 1000000
> iterations by default.  Later I added the DO_TEST that lets you test
> the empty loop just because I was curious.
> 
> A more realistic test with the two 'mov' instructions inside the loops
> still only takes 16 clocks, so I'm wondering why you get 60?  Does the
> vsyscall add that much overhead?  With this I get 29-30 clocks per loop
> on Pentium II:

This is the x86-64 test code I used. It's basically an emulation of the vsyscall
(including indirect call) in user space.

rdtscp shows less cycles, so it's not all overhead of the infrastructure.

-Andi


K8 E stepping: 

getpid 168 cycles
vgetcpu lsl 79 cycles
vgetcpu cached 15 cycles

K8 F stepping:

getpid 162 cycles
vgetcpu lsl 77 cycles
vgetcpu rdtscp 32 cycles
vgetcpu cached 15 cycles

Nocona: 

getpid 1491 cycles
vgetcpu lsl 130 cycles
vgetcpu cached 26 cycles