Re: remove BufferBlockPointers for speed and space - Mailing list pgsql-patches

FreeBSD 5.4 RELEASE gcc 3.4.2 on Intel dual PIII 1Ghz

[postgres:~/develop/c/testbuf]$ gcc -Wall -o testbuf testbuf.c
[postgres:~/develop/c/testbuf]$ ./testbuf
duration round 1 of array method: 1.737 ms
duration round 2 of array method: 1.676 ms
duration round 3 of array method: 1.527 ms
duration round 1 of mul method: 0.548 ms
duration round 2 of mul method: 0.548 ms
duration round 3 of mul method: 0.546 ms
duration round 1 of shift method: 0.593 ms
duration round 2 of shift method: 0.592 ms
duration round 3 of shift method: 0.575 ms
[postgres:~/develop/c/testbuf]$ gcc -Wall -O2 -o testbuf testbuf.c
[postgres:~/develop/c/testbuf]$ ./testbuf
duration round 1 of array method: 0.169 ms
duration round 2 of array method: 0.165 ms
duration round 3 of array method: 0.165 ms
duration round 1 of mul method: 0.164 ms
duration round 2 of mul method: 0.164 ms
duration round 3 of mul method: 0.165 ms
duration round 1 of shift method: 0.164 ms
duration round 2 of shift method: 0.164 ms
duration round 3 of shift method: 0.165 ms

Looks to me like -O2 makes the difference very small (on this
platform/gcc combo) - is 5/169 worth doing?

BTW - I patched the program to stop gcc whining:

--- testbuf.c.orig      Thu Aug 11 18:41:29 2005
+++ testbuf.c   Thu Aug 11 18:37:43 2005
@@ -2,6 +2,7 @@
   * testbuf.c
   */
  #include <stdio.h>
+#include <stdlib.h>
  #include <sys/file.h>
  #include <sys/param.h>
  #include <sys/stat.h>
@@ -66,6 +67,8 @@
         (long) (stop_t.tv_usec - start_t.tv_usec) % 1000);
    }
   }
+
+ return 0;
  }


Qingqing Zhou wrote:
> "Tom Lane" <tgl@sss.pgh.pa.us> writes:
>
>>Also, I would like to see the actual test code.  I wonder whether what
>>you measured is the ability of the compiler to optimize references to
>>successive elements of an array inside a loop; that has little or
>>nothing to do with the typical usage of BufferGetBlock().
>>
>
>
> The source code is attached.
>
> compiled with "gcc testbuf.c". I tried -O2 actually, and it turns out that
> the timing is reduced a lot so not believable.
> ---
>
> /*
>  * testbuf.c
>  */
> #include <stdio.h>
> #include <sys/file.h>
> #include <sys/param.h>
> #include <sys/stat.h>
> #include <sys/time.h>
> #include <unistd.h>
> #include <fcntl.h>
>
> #define BLCKSZ  8192
> #define NBuffers 80000
>
> typedef void* Block;
>
> int main(void)
> {
>  int  i, round, method;
>  Block k, start;
>  struct timeval start_t, stop_t;
>  long usecs;
>  Block *array = (Block *) calloc(NBuffers, sizeof(Block));
>
>  start = (Block)0xff3386;
>  for (i = 0; i < NBuffers; i++)
>   array[i] = start + BLCKSZ*i;
>
>  for (method = 0; method < 3; method ++)
>  {
>   start = (Block)0xff3386;
>   for (round = 0; round < 3; round ++)
>   {
>    gettimeofday(&start_t, NULL);
>    if (method == 0)
>    {
>     for (i = 0; i < NBuffers; i++)
>      k = array[i];
>    }
>    if (method == 1)
>    {
>     for (i = 0; i < NBuffers; i++)
>      k = start + i*BLCKSZ;
>    }
>    if (method == 2)
>    {
>     for (i = 0; i < NBuffers; i++)
>      k = start + (i<<13);
>    }
>    gettimeofday(&stop_t, NULL);
>
>    if (stop_t.tv_usec < start_t.tv_usec)
>    {
>      stop_t.tv_sec--;
>      stop_t.tv_usec += 1000000;
>    }
>
>    usecs = (long) (stop_t.tv_sec - start_t.tv_sec) * 1000000
>      + (long) (stop_t.tv_usec - start_t.tv_usec);
>
>    fprintf (stdout, "duration round %d of %s method: %ld.%03ld ms\n",
>        round + 1,
>        method==0?"array":method==1?"mul":"shift",
>        (long) ((stop_t.tv_sec - start_t.tv_sec) * 1000 +
>          (stop_t.tv_usec - start_t.tv_usec) / 1000),
>        (long) (stop_t.tv_usec - start_t.tv_usec) % 1000);
>   }
>  }
> }
>
>
>
> ---------------------------(end of broadcast)---------------------------
> TIP 9: In versions below 8.0, the planner will ignore your desire to
>        choose an index scan if your joining column's datatypes do not
>        match
>
>