(2014/01/30 8:29), Tom Lane wrote:
> Andrew Dunstan <andrew@dunslane.net> writes:
>> I could live with just stddev. Not sure others would be so happy.
>
> FWIW, I'd vote for just stddev, on the basis that min/max appear to add
> more to the counter update time than stddev does; you've got
> this:
>
> + e->counters.total_sqtime += total_time * total_time;
>
> versus this:
>
> + if (e->counters.min_time > total_time || e->counters.min_time == EXEC_TIME_INIT)
> + e->counters.min_time = total_time;
> + if (e->counters.max_time < total_time)
> + e->counters.max_time = total_time;
>
> I think on most modern machines, a float multiply-and-add is pretty
> darn cheap; a branch that might or might not be taken, OTOH, is a
> performance bottleneck.
>
> Similarly, the shared memory footprint hit is more: two new doubles
> for min/max versus one for total_sqtime (assuming we're happy with
> the naive stddev calculation).
>
> If we felt that min/max were of similar value to stddev then this
> would be mere nitpicking. But since people seem to agree they're
> worth less, I'm thinking the cost/benefit ratio isn't there.
Why do you surplus worried about cost in my patch? Were three double variables
assumed a lot of memory, or having lot of calculating cost? My test result
showed LWlock bottele-neck is urban legend. If you have more fair test
pattern, please tell me, I'll do it if the community will do fair judge.
Regards,
--
Mitsumasa KONDO
NTT Open Source Software Center