On Sun, Sep 22, 2013 at 08:46:55PM +0200, Fabien wrote:
> pgbench: reduce and compensate throttling underestimation bias.
>
> This is a consequence of relying on an integer random generator,
> which allow to ensure that delays inserted stay reasonably in
> range of the target average delay.
>
> The bias was about 0.5% with 1000 distinct values. Multiplier added
> to compensate the 0.05% bias with 10000 distinct integer values,
> so there is no bias now.
> --- a/contrib/pgbench/pgbench.c
> +++ b/contrib/pgbench/pgbench.c
> @@ -929,13 +929,17 @@ top:
> * that the series of delays will approximate a Poisson distribution
> * centered on the throttle_delay time.
> *
> - * 1000 implies a 6.9 (-log(1/1000)) to 0.0 (log 1.0) delay multiplier.
> + * 10000 implies a 9.2 (-log(1/10000)) to 0.0 (log 1) delay multiplier,
> + * and results in a 0.055 % target underestimation bias:
> + *
> + * SELECT 1.0/AVG(-LN(i/10000.0)) FROM generate_series(1,10000) AS i;
> + * = 1.000552717032611116335474
> *
> * If transactions are too slow or a given wait is shorter than
> * a transaction, the next transaction will start right away.
> */
> - int64 wait = (int64)
> - throttle_delay * -log(getrand(thread, 1, 1000)/1000.0);
> + int64 wait = (int64) (throttle_delay *
> + 1.00055271703 * -log(getrand(thread, 1, 10000)/10000.0));
>
> thread->throttle_trigger += wait;
>
On Sun, Sep 22, 2013 at 10:08:54AM +0200, Fabien COELHO wrote:
>> Then I understand why you want to remove the bias, but
>> why do you also increase the upper bound?
>
> Because the bias was significantly larger for 1000 (about 0.5%), so this
> also contributed to reduce said bias, and 9.2 times the average target
> seems as reasonnable a bound as 6.9.
The magnitude of the bias is unimportant if known exactly, as you do here.
Smaller quanta do give you more possible distinct sleep durations, which is a
nice bonus. What I'm hearing is that you don't especially want to change the
upper bound, but changing it was an acceptable side effect. However, we can
meet all those goals simultaneously with simpler code, can we not?
int64 wait = (int64) (throttle_delay * Min(7.0, -log(1 - pg_erand48(thread->random_state))));
--
Noah Misch
EnterpriseDB http://www.enterprisedb.com