Flexible I/O Tester development
 help / color / mirror / Atom feed
* [PATCH] gettime: use nsec in get_cycles_per_msec division
@ 2018-11-29 18:14 vincentfu
  2018-11-29 18:21 ` Jens Axboe
  0 siblings, 1 reply; 3+ messages in thread
From: vincentfu @ 2018-11-29 18:14 UTC (permalink / raw)
  To: axboe, fio; +Cc: Vincent Fu

From: Vincent Fu <vincent.fu@wdc.com>

Since we now have ntime_since() we can use nsec in the division for
get_cycles_per_msec(). This makes the cycles_per_msec value that fio
ultimately uses more stable since fio will no longer be using values
truncated to usec.

While we're here also modify some debug prints to make it explicit that
fio ultimately uses a trimmed mean in its time calculations.

On platforms where only gettimeofday() is available this will be no
worse than the original code since we will revert back to using times
truncated to usec.

With the patch, the last three digits of the trimmed mean have a tight
range from 784-786. Without the patch, the corresponding value (labeled
avg) has a much wider range from 821-958. Notice also that the standard
error S is an order of magnitude smaller with the patch.

*** WITH PATCH ***
$ ./fio --debug=time --cpuclock-test | grep mean
time     315   min=3466785, max=3466796, mean=3466789.120000, S=0.057234, N=50
time     315   trimmed mean=3466789, N=29
$ ./fio --debug=time --cpuclock-test | grep mean
time     329   min=3466784, max=3466795, mean=3466789.480000, S=0.053641, N=50
time     329   trimmed mean=3466788, N=36
$ ./fio --debug=time --cpuclock-test | grep mean
time     343   min=3466786, max=3466794, mean=3466789.220000, S=0.036654, N=50
time     343   trimmed mean=3466788, N=42
$ ./fio --debug=time --cpuclock-test | grep mean
time     357   min=3466785, max=3466794, mean=3466789.080000, S=0.053580, N=50
time     357   trimmed mean=3466788, N=33
$ ./fio --debug=time --cpuclock-test | grep mean
time     371   min=3466785, max=3466794, mean=3466789.600000, S=0.043519, N=50
time     371   trimmed mean=3466789, N=36
$ ./fio --debug=time --cpuclock-test | grep mean
time     385   min=3466785, max=3466794, mean=3466789.280000, S=0.052534, N=50
time     385   trimmed mean=3466789, N=32
$ ./fio --debug=time --cpuclock-test | grep mean
time     407   min=3466786, max=3466796, mean=3466789.520000, S=0.042616, N=50
time     407   trimmed mean=3466789, N=41
$ ./fio --debug=time --cpuclock-test | grep mean
time     513   min=3466785, max=3466796, mean=3466789.220000, S=0.051316, N=50
time     513   trimmed mean=3466789, N=35

*** WIHTOUT PATCH ***
$ ./fio-3.11 --debug=time --cpuclock-test | grep 'mean\|avg'
time     439   avg: 3466821
time     439   min=3466790, max=3467006, mean=3466838.060000, S=0.959748
$ ./fio-3.11 --debug=time --cpuclock-test | grep 'mean\|avg'
time     455   avg: 3466958
time     455   min=3466796, max=3466996, mean=3466945.880000, S=0.789178
$ ./fio-3.11 --debug=time --cpuclock-test | grep 'mean\|avg'
time     469   avg: 3466887
time     469   min=3466790, max=3466984, mean=3466887.500000, S=0.825069
$ ./fio-3.11 --debug=time --cpuclock-test | grep 'mean\|avg'
time     485   avg: 3466951
time     485   min=3466818, max=3466996, mean=3466939.420000, S=0.873334
$ ./fio-3.11 --debug=time --cpuclock-test | grep 'mean\|avg'
time     499   avg: 3466957
time     499   min=3466793, max=3467009, mean=3466950.140000, S=0.764626
$ ./fio-3.11 --debug=time --cpuclock-test | grep 'mean\|avg'
time     608   avg: 3466954
time     608   min=3466818, max=3466981, mean=3466937.040000, S=0.862232
---
 gettime.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/gettime.c b/gettime.c
index 7702193d..272a3e62 100644
--- a/gettime.c
+++ b/gettime.c
@@ -239,13 +239,13 @@ static unsigned long get_cycles_per_msec(void)
 		__fio_gettime(&e);
 		c_e = get_cpu_clock();
 
-		elapsed = utime_since(&s, &e);
-		if (elapsed >= 1280)
+		elapsed = ntime_since(&s, &e);
+		if (elapsed >= 1280000)
 			break;
 	} while (1);
 
 	fio_clock_source = old_cs;
-	return (c_e - c_s) * 1000 / elapsed;
+	return (c_e - c_s) * 1000000 / elapsed;
 }
 
 #define NR_TIME_ITERS	50
@@ -298,10 +298,10 @@ static int calibrate_cpu_clock(void)
 
 	avg /= samples;
 	cycles_per_msec = avg;
-	dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
-	dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
+	dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f, N=%d\n",
 			(unsigned long long) minc,
-			(unsigned long long) maxc, mean, S);
+			(unsigned long long) maxc, mean, S, NR_TIME_ITERS);
+	dprint(FD_TIME, "trimmed mean=%llu, N=%d\n", (unsigned long long) avg, samples);
 
 	max_ticks = MAX_CLOCK_SEC * cycles_per_msec * 1000ULL;
 	max_mult = ULLONG_MAX / max_ticks;
-- 
2.17.1



^ permalink raw reply related	[flat|nested] 3+ messages in thread

* Re: [PATCH] gettime: use nsec in get_cycles_per_msec division
  2018-11-29 18:14 [PATCH] gettime: use nsec in get_cycles_per_msec division vincentfu
@ 2018-11-29 18:21 ` Jens Axboe
  2018-11-29 18:31   ` Vincent Fu
  0 siblings, 1 reply; 3+ messages in thread
From: Jens Axboe @ 2018-11-29 18:21 UTC (permalink / raw)
  To: vincentfu, fio; +Cc: Vincent Fu

On 11/29/18 11:14 AM, vincentfu@gmail.com wrote:
> From: Vincent Fu <vincent.fu@wdc.com>
> 
> Since we now have ntime_since() we can use nsec in the division for
> get_cycles_per_msec(). This makes the cycles_per_msec value that fio
> ultimately uses more stable since fio will no longer be using values
> truncated to usec.
> 
> While we're here also modify some debug prints to make it explicit that
> fio ultimately uses a trimmed mean in its time calculations.
> 
> On platforms where only gettimeofday() is available this will be no
> worse than the original code since we will revert back to using times
> truncated to usec.
> 
> With the patch, the last three digits of the trimmed mean have a tight
> range from 784-786. Without the patch, the corresponding value (labeled
> avg) has a much wider range from 821-958. Notice also that the standard
> error S is an order of magnitude smaller with the patch.

This is great, thanks Vincent. Applied.

-- 
Jens Axboe



^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH] gettime: use nsec in get_cycles_per_msec division
  2018-11-29 18:21 ` Jens Axboe
@ 2018-11-29 18:31   ` Vincent Fu
  0 siblings, 0 replies; 3+ messages in thread
From: Vincent Fu @ 2018-11-29 18:31 UTC (permalink / raw)
  To: Jens Axboe, vincentfu@gmail.com, fio@vger.kernel.org

On 11/29/18 1:21 PM, Jens Axboe wrote:
> On 11/29/18 11:14 AM, vincentfu@gmail.com wrote:
>> From: Vincent Fu <vincent.fu@wdc.com>
>>
>> Since we now have ntime_since() we can use nsec in the division for
>> get_cycles_per_msec(). This makes the cycles_per_msec value that fio
>> ultimately uses more stable since fio will no longer be using values
>> truncated to usec.
>>
>> While we're here also modify some debug prints to make it explicit that
>> fio ultimately uses a trimmed mean in its time calculations.
>>
>> On platforms where only gettimeofday() is available this will be no
>> worse than the original code since we will revert back to using times
>> truncated to usec.
>>
>> With the patch, the last three digits of the trimmed mean have a tight
>> range from 784-786. Without the patch, the corresponding value (labeled
>> avg) has a much wider range from 821-958. Notice also that the standard
>> error S is an order of magnitude smaller with the patch.
> This is great, thanks Vincent. Applied.
I made a mistake in the commit message. Instead of "a tight range from
784-786" it should be 788-789.


^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2018-11-29 18:31 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2018-11-29 18:14 [PATCH] gettime: use nsec in get_cycles_per_msec division vincentfu
2018-11-29 18:21 ` Jens Axboe
2018-11-29 18:31   ` Vincent Fu

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox