On Thu, 24 Nov 2005, Qingqing Zhou wrote:
>
> I may need to write some separate tests to see if this is what we should
> pay for bus lock instruction.
>
Here I come up with a test program to see how spinlock costs:
$/pgsql/src/backend/storage/lmgr#./a.out
Spinlock pair(2648542) duration: 143.134 ms
$/pgsql/src/backend/storage/lmgr#./a.out
Spinlock pair(2648542) duration: 143.107 ms
$/pgsql/src/backend/storage/lmgr#./a.out
Spinlock pair(2648542) duration: 143.104 ms
So seems lock instruction really costs ...
Regards,
Qingqing
---
/*
* spintest.c -
* Test spinlock acquire/release without concurrency.
*
* To compile (the -pg is to match the gprof make I used):
* backend/storage/lmgr#gcc -O2 -pg -Wall -I ../../../include/
spintest.c
*/
#include "postgres.h"
#include "storage/lwlock.h"
#include "storage/spin.h"
#include <sys/time.h>
#define TIMES 2648542
int NumLocks = 0;
void
s_lock(volatile slock_t *lock, const char *file, int line)
{
fprintf(stderr, "should never be here\n");
abort();
}
int
main(void)
{
int i;
slock_t lock = 0;
struct timeval start_t, stop_t;
long usecs;
gettimeofday(&start_t, NULL);
for (i = 0; i < TIMES; i ++)
{
SpinLockAcquire_NoHoldoff(&lock);
/* pretend to do something */
NumLocks ++;
SpinLockRelease_NoHoldoff(&lock);
}
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, "Spinlock pair(%u) duration: %ld.%03ld ms\n",
TIMES,
(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);
return 0;
}
---------------------------(end of broadcast)---------------------------
TIP 4: Have you searched our list archives?
http://archives.postgresql.org
