On 17/02/15 16:12, Andres Freund wrote:
> On 2015-02-17 15:50:39 +0100, Petr Jelinek wrote:
>> On 17/02/15 03:07, Petr Jelinek wrote:
>>> On 17/02/15 03:03, Andrew Dunstan wrote:
>>>> On 02/16/2015 08:57 PM, Andrew Dunstan wrote:
>>>>>> Average of 3 runs of read-only pgbench on my system all with
>>>>>> pg_stat_statement activated:
>>>>>> HEAD: 20631
>>>>>> SQRT: 20533
>>>>>> SQRTD: 20592
>
>>>>> So using sqrtd the cost is 0.18%. I think that's acceptable.
>
>>>> Actually, sqrt/sqrtd is not called in accumulating the stats, only in
>>>> the reporting function. So it looks like the difference here should be
>>>> noise. Maybe we need some longer runs.
>
>>> Yes there are variations between individual runs so it might be really
>>> just that, I can leave it running for much longer time tomorrow.
>
>> Ok so I let it run for more than hour on a different system, the difference
>> is negligible - 14461 vs 14448 TPS. I think there is bigger difference
>> between individual runs than between the two versions...
>
> These numbers sound like you measured them without concurrency, am I
> right? If so, the benchmark isn't that interesting - the computation
> happens while a spinlock is held, and that'll mainly matter if there are
> many backends running at the same time.
>
It's pgbench -j16 -c32
-- Petr Jelinek http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training &
Services
