Timers and timing, was: MySQL Performance 6.0rc1

[Robert Watson]

On Fri, 28 Oct 2005, David Xu wrote:

>>> On the other hand, a lower risk change might be to simply add a new 
>>> CLOCK_ type for lower resolution, and have a timer synchronize a 
>>> variable to the system clock once every 1/10 of a second.  This avoids 
>>> having to muck with VM layout, etc.
>> 
>> Is the CLOCK_* namespace ours to muck about with in the first place ?
>> 
> I prefer this way, can you implement it? The global page idea is a 
> complex, someone can slowly work on it, there are many things can be 
> done, for example, fast syscall using sysenter/sysexit.

Just as an experiment, I added two new time counters:

CLOCK_POOR - use a cached time returned by getnanotime(), which may be out
              of date by up to 10/hz (see phk's e-mail for the correct
              number).

CLOCK_SECOND - Same as CLOCK_POOR, only truncate the nanoseconds field to
                0, and return 1 second as the resolution via
                clock_getres().

Patch is attached.

Performance measurements on a P4 Xeon, dual processor with UP and SMP 
kernel configurations, no debugging.  Measured using 10,000 loops of each 
system call, 12 samples per clock type.  Measurements are in nanoseconds, 
measured using CLOCK_REALTIME, which has a resolution of 280ns reported 
via clock_getres().  Not surprisingly, the two above clocks measure about 
the same (they should do), and are a lot faster than real time 
measurements (they should be).  As currently implemented, gettimeofday() 
appears to cost the same as CLOCK_REALTIME.

It would be interesting to distribution of clock wrongness was, but I'm 
not set up to measure that currently.

Robert N M Watson

x 7UP/gettimeofday
+ 7UP/clock_gettime_poor
* 7UP/clock_gettime_realtime
% 7UP/clock_gettime_second
+--------------------------------------------------------------------------+
|       @                                                         *        |
|       @                                                         *        |
|       @                                                         *        |
|       @                                                         *        |
|+      @                                                         *        |
|@      @                                                        x*        |
|@%@   @@                                                   x*   ***      *|
|  |__A_M|                                                     |_AA|_|     |
+--------------------------------------------------------------------------+
     N           Min           Max        Median           Avg        Stddev
x  12          1410          1507          1502       1493.75     26.574509
+  12           523           634           624     595.58333     47.586206
Difference at 95.0% confidence
         -898.167 +/- 32.632
         -60.1283% +/- 2.18457%
         (Student's t, pooled s = 38.5399)
*  12          1437          1627          1503     1507.6667     42.401401
No difference proven at 95.0% confidence
%  12           524           631           623           595     45.696628
Difference at 95.0% confidence
         -898.75 +/- 31.6491
         -60.1674% +/- 2.11877%
         (Student's t, pooled s = 37.379)



x 7SMP/gettimeofday
+ 7SMP/clock_gettime_poor
* 7SMP/clock_gettime_realtime
% 7SMP/clock_gettime_second
+--------------------------------------------------------------------------+
|@   @                                                                 * x*|
|@   @                                                                 * x*|
|@%  @%                                                                * x*|
|@@  @@                                                               **x**|
||MA_M                                                                 |AMM|
+--------------------------------------------------------------------------+
     N           Min           Max        Median           Avg        Stddev
x  12          1406          1443          1436     1425.4167     16.527984
+  12           503           566           517     527.08333     26.352362
Difference at 95.0% confidence
         -898.333 +/- 18.6239
         -63.0225% +/- 1.30656%
         (Student's t, pooled s = 21.9957)
*  12          1404          1450          1444     1430.0833     19.327128
No difference proven at 95.0% confidence
%  12           505           562           549         531.5     26.182923
Difference at 95.0% confidence
         -893.917 +/- 18.538
         -62.7127% +/- 1.30054%
         (Student's t, pooled s = 21.8943)
-------------- next part --------------
--- //depot/vendor/freebsd/src/sys/kern/kern_time.c	2005/10/23 23:00:53
+++ //depot/user/rwatson/clock/src/sys/kern/kern_time.c	2005/10/28 17:42:44
@@ -225,6 +225,14 @@
 	case CLOCK_MONOTONIC:
 		nanouptime(ats);
 		break;
+	case CLOCK_SECOND:
+		getnanotime(ats);
+		ats->tv_nsec = 0;
+		break;
+	case CLOCK_POOR:
+		getnanotime(ats);
+		ats->tv_nsec = 0;
+		break;
 	default:
 		return (EINVAL);
 	}
@@ -306,6 +314,7 @@
 	switch (clock_id) {
 	case CLOCK_REALTIME:
 	case CLOCK_MONOTONIC:
+	case CLOCK_POOR:
 		/*
 		 * Round up the result of the division cheaply by adding 1.
 		 * Rounding up is especially important if rounding down
@@ -318,6 +327,10 @@
 		/* Accurately round up here because we can do so cheaply. */
 		ts->tv_nsec = (1000000000 + hz - 1) / hz;
 		break;
+	case CLOCK_SECOND:
+		ts->tv_sec = 1;
+		ts->tv_nsec = 0;
+		break;
 	default:
 		return (EINVAL);
 	}
--- //depot/vendor/freebsd/src/sys/sys/time.h	2005/04/02 12:35:19
+++ //depot/user/rwatson/clock/src/sys/sys/time.h	2005/10/28 17:42:44
@@ -238,6 +238,8 @@
 #define CLOCK_VIRTUAL	1
 #define CLOCK_PROF	2
 #define CLOCK_MONOTONIC	4
+#define	CLOCK_SECOND	5
+#define	CLOCK_POOR	6
 #endif
 
 #ifndef TIMER_ABSTIME