system time - windows specific problem

On 18/07/2008 5:26 PM, Christophe Dutang wrote:

Hi all,
I'm currently implementing quasi random generation (torus) on R   
(package randtoolbox available on CRAN). Even if it is not a great   
idea, I decided to use the machine time to initiate the seed. So  
when  the seed is not specified by the user, the pkg uses the time  
machine  time.
Hence the following R code should produce different uniform  
variates :

for(i in 1:10)print(torus(1))

But on windows, I get :

library(randtoolbox)
for(i in 1:10)print(torus(1))

[1] 0.1313116
[1] 0.1313116
[1] 0.1313116
[1] 0.1313116
[1] 0.1313116
[1] 0.1313116
[1] 0.1313116
[1] 0.1313116
[1] 0.1313116
[1] 0.1313116

while on mac os, we get :

library(randtoolbox)
for(i in 1:10)print(torus(1))

[1] 0.9730577
[1] 0.9124289
[1] 0.5123534
[1] 0.4640765
[1] 0.9367557
[1] 0.2945414
[1] 0.3671455
[1] 0.2698379
[1] 0.6036739
[1] 0.6762776

I think I know where the problem is. in the C file, randtoolbox.c,  
I  use the machine time as follows
void randSetSeed()
{
        /* struct timeval {
                unsigned long tv_sec; // seconds since Jan. 1, 1970
                long tv_usec; // and microseconds  };
        */
        struct timeval tv;
        //take the machine time
        gettimeofday (&tv, NULL);
        // [ 2^16 * microsecond ] xor [ second ]
        seed = ((unsigned long long) tv.tv_usec << 16) ^ tv.tv_sec;
        isInit = 1;
}
and the C function randSetSeed is called. (in the header file, I   
include <sys/time.h>).
I think on windows there is no micro second precision and so the   
"randomize" seed is the same between two consecutive calls.

The same will eventually be true on other machines, when they are  
fast enough to evaluate two calls during the same microsecond.  (I'd  
guess they're already fast enough if you put that loop in C code  
instead of R code.)  But why would this ever happen?  Surely your  
code doesn't make this call for every random number call; like R, it  
should make it at most once per session.

There is a higher resolution counter in Windows (see  
QueryPerformanceCounter) if you really need it, and you may have  
access to the hardware cycle counter if you need even better  
resolution.  But I don't see why you would need these.

Duncan Murdoch

I try many things to deal with this problem (for example include   
<windows.h> and use windows specific time function), but it does  
not  work. I'm asking for help on R mailing list because what solve  
on R  for the runif function.
You could answer me to go to the R source. Of course, I look there  
my  code is based on this code. But in the RNG.c file which  
implements  random number generation. they use line 271-289
#include <time.h>
#ifdef HAVE_SYS_TIME_H
# include <sys/time.h>
#endif
static void Randomize(RNGtype kind)
{
/* Only called by  GetRNGstate() when there is no .Random.seed */
    Int32 seed;
#if HAVE_GETTIMEOFDAY
  {
    struct timeval tv;
    gettimeofday (&tv, NULL);
    seed = ((uint64_t) tv.tv_usec << 16) ^ tv.tv_sec;
  }
#elif HAVE_TIME
    seed = (Int32) time(NULL);
#else
    /* unlikely, but use random contents */
#endif
    srand(seed);
    RNG_Init(kind, seed);
}
If I try to use directly the time function on windows rather than   
gettimeofday (which should not be on windows) but it does not  
solve  the problem.
I'm wondering how R has solved this issue?
That's where I'm asking for any help.
Thanks in advance
Christophe
PS : I use in both examples R 2.7.1 first on windows XP, second on  
mac  os leopard.