This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: advise on future of glibc (mktime issue)

On Tue, Jun 07, 2005 at 10:35:48AM -0400, wrote:
>  I contacted RH and they say they will not change the behavior of their 
> glibc because they prefer "correctness over performance" and they 
> believe glibc should be checking for changes to the timezone. Their 
> suggestions are: build our own glibc and loose RH support, rewrite 
> mktime for our own use (I tried that but seemed a little complicated and 
> looses the benefit of the library per se), and lastly they suggested to 
> set TZ=/etc/localtime (which makes the program run a little faster, but 
> not as fast as RH7.3).

I certainly can't reproduce contemporary glibc being any slower than
RH7.3 glibc with TZ=/etc/localtime:

$ LC_ALL=C time ~/7.3-glibc/ --library-path ~/7.3-glibc/ /tmp/mktimetest
1.74user 0.00system 0:01.74elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+93minor)pagefaults 0swaps
$ LC_ALL=C TZ=/etc/localtime time ~/7.3-glibc/ --library-path ~/7.3-glibc/ /tmp/mktimetest
1.66user 0.00system 0:01.66elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+93minor)pagefaults 0swaps
$ LC_ALL=C time /lib/ --library-path /lib /tmp/mktimetest
1.94user 13.73system 0:15.68elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+100minor)pagefaults 0swaps
$ TZ=/etc/localtime time /lib/ --library-path /lib /tmp/mktimetest
1.36user 0.00system 0:01.36elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+100minor)pagefaults 0swaps

~/7.3-glibc/ has glibc-2.2.5-34.i686.rpm's /lib dir unpacked, /lib
contains 2.3.5 glibc.  All is on a FC3 box, but with TZ=/etc/localtime the
loop does not do any syscalls, so the kernel shouldn't matter much.
Here, glibc 2.3.5 is actually faster than 2.2.5 if TZ=/etc/localtime is

>  It is widely known that FC1,2,3 and RHEL* are slow systems, and I 
> always figured it was for the better, then I question, how may of this 

Do you have any data to back up your claims?

> little changes in glibc exists in RH's patches that by themselves may 
> not be a big performance hit, but put them together and you get a slower 
> system ?

You can easily check yourself, each Red Hat glibc src.rpm includes
a tarball with the upstream CVS checkout and a patch.  I'm not aware of
any changes from upstream glibc that could noticeably affect performance.

>  I need to know the weekday for millions of dates, I guess I could use
> a hash (I would need to find a library for that) for it, but everything
> was fine until I started using RHEL4 with the patched glibc. Any
> suggestions ?

I might be missing something obvious, but if all you need is the weekday
for dates stored in the year*10000+month*100+day format and you need
to do it that many times, then surely writing your own function
will be uncomparably faster than doing expensive mktime and localtime.

I haven't put any effort into optimizing this and you can get away
even without mktime calls or just one, not just caching tm_wday
of Jan, 1st for each year.  Still, it is more than 60 times faster:

LC_ALL=C time /lib/ --library-path /lib /tmp/mktimetest2
0.02user 0.00system 0:00.02elapsed 96%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+101minor)pagefaults 0swaps

int real_get_dayofweek(int date)
 // extract day/yr/mn
 int day = date % 100;
 int month = (date % 10000) / 100;
 int year = date / 10000;

 struct tm time;
 time.tm_sec = 0;  time.tm_min = 30;  time.tm_hour = 9;  time.tm_mday = day;  time.tm_mon = month - 1;
 time.tm_year = year - 1900;

 time_t time1 = mktime(&time);

 return(localtime(&(time1))->tm_wday); // Sun=0, Sat=6

} //get_dayofweek

static unsigned char year_to_wday[200];

int get_dayofweek(int date)
  // extract day/yr/mn
  int day = date % 100;
  int month = (date % 10000) / 100;
  int year = date / 10000;
  int leapyear = (year & 3) == 0 && ((year % 100) != 0 || (year % 400) == 0);
  static const unsigned short int mon_yday[2][13] =
    /* Normal years.  */
    { 0, 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334 },
    /* Leap years.  */
    { 0, 0, 31, 60, 91, 121, 152, 182, 213, 244, 274, 305, 335 }
  if (year < 1900 || year >= 2100)
    return real_get_dayofweek (date);
  return (year_to_wday[year - 1900] + mon_yday[leapyear][month] + day - 1) % 7;

void init_year_to_wday (void)
  int date;
  for (date = 0; date < 200; date++)
    year_to_wday[date] = real_get_dayofweek(date * 10000 + 19000101);


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]