Re: [Patch] Step2: don't use _M_out_lim in filebuf, have filebuf non unified, etc...

[Nathan Myers <ncm-nospam at cantrip dot org>]

On Tue, Jun 24, 2003 at 06:33:33PM +0200, Paolo Carlini wrote:
> >On Tue, Jun 24, 2003 at 12:17:26AM +0200, Paolo Carlini wrote:
> Nathan Myers wrote:
> >Once we get the streambuf inlines shuffled around between
> >std/std_streambuf.h and bits/streambuf.tcc, we'll be ready to run
> >some races.  Eat our dust, stdio!  :-)
> 
> Anyway, *right now*, on my P4, for 9876
> 
> 1.990u 0.370s 0:02.37 99.5%     0+0k 0+0io 215pf+0w   (stdio unlocked)
> 1.480u 0.380s 0:01.94 95.8%     0+0k 0+0io 216pf+0w   (iostream)

I ran another benchmark, attached below, on several compilers
on P3 and Opteron.  All builds are -O2.

P3 1.4GHz.
----------
gcc version 2.95.4 20011002 (Debian prerelease)
stdio (unlocked):       2.54u   0.10s   4.60r
streambuf:              21.87u  0.13s   23.24r

gcc version 3.2.2 20021230 (prerelease)
stdio (unlocked):       1.88u   0.07s   2.52r
streambuf:              14.34u  0.06s   14.87r

Mainline gcc version 3.4 20030624 (experimental)
stdio (unlocked):       1.82u   0.13s   3.57r
streambuf:              5.89u   0.05s   7.45r

Mainline w/ inlines
stdio (unlocked):       1.85u   0.10s   2.65r
streambuf:              2.10u   0.09s   2.79r


Opteron  1.4GHz
---------------
gcc 3.2.2
stdio (unlocked):       1.60u   1.14s   2.93r
streambuf:              9.56u   0.98s   12.22r

gcc version 3.4 20030613:
stdio (unlocked):       1.69u   1.02s   4.48r
streambuf:              4.81u   0.65s   6.65r

Mainline gcc version 3.4 20030624 (experimental)
stdio (unlocked):       1.68u   0.28s   2.52r
streambuf:              4.48u   0.40s   5.37r

Mainline w/inlines
stdio (unlocked):       1.64u   0.34s   3.52r
streambuf:              3.08u   0.33s   4.91r

$ time dd if=/dev/zero of=/dev/null bs=8k count=32768
real    0m0.647s
user    0m0.000s
sys     0m0.650s

We still have some room for improvement, perhaps. :-)

Nathan Myers
ncm-nospam@cantrip.org

p.s.  This is the test program.  It exercises both input and output, and 
uses /dev/zero and /dev/null (respectively) to minimize OS effects.

=========
#include <iostream>
#include <fstream>
#include <stdio.h>
#include <sys/times.h>
#include <sys/time.h>
#include <time.h>
#include <unistd.h>

const unsigned long count = 1ul << 28;

int main()
{
  const long ticksper = sysconf(_SC_CLK_TCK);

  tms before;
  times(&before);
  timeval tvbefore;
  gettimeofday(&tvbefore, 0);

  FILE* fpi = fopen("/dev/zero", "r");
  FILE* fpo = fopen("/dev/null", "w");
  for (unsigned long i = 0; i < count; ++i) {
    int c = getc_unlocked(fpi); 
    if (c != EOF) {
      putc_unlocked(c, fpo);
    }
  }
  fclose(fpi);
  fclose(fpo);

  tms middle;
  times(&middle);
  timeval tvmiddle;
  gettimeofday(&tvmiddle, 0);

  std::filebuf  in;
  in.open("/dev/zero", std::ios::in);
  std::filebuf out;
  out.open("/dev/null", std::ios::out);
  for (unsigned long i = 0; i < count; ++i) {
    int c = in.sbumpc(); 
    if (c != EOF) {
      out.sputc(c);
    }
  }
  in.close();
  out.close();

  tms after;
  times(&after);
  timeval tvafter;
  gettimeofday(&tvafter, 0);

  printf("stdio (unlocked):\t%.02fu\t%.02fs\t%.02fr\n", 
      double(middle.tms_utime - before.tms_utime) / ticksper,
      double(middle.tms_stime - before.tms_stime) / ticksper,
      double(tvmiddle.tv_sec) + .000001 * double(tvmiddle.tv_usec) -
      double(tvbefore.tv_sec) + .000001 * double(tvbefore.tv_usec)
    );
  printf("streambuf:       \t%.02fu\t%.02fs\t%.02fr\n", 
      double(after.tms_utime - middle.tms_utime) / ticksper,
      double(after.tms_stime - middle.tms_stime) / ticksper,
      double( tvafter.tv_sec) + .000001 * double( tvafter.tv_usec) -
      double(tvmiddle.tv_sec) + .000001 * double(tvmiddle.tv_usec)
      );

  return 0;
}