KONDO Mitsumasa <kondo.mitsumasa@lab.ntt.co.jp> writes:
> (2014/01/30 8:29), Tom Lane wrote:
>> 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.
[ shrug... ] Standard pgbench tests are useless for measuring microscopic
issues like this one; whatever incremental cost is there will be swamped
by client communication and parse/plan overhead. You may have proven that
the overhead isn't 10% of round-trip query time, but we knew that without
testing.
If I were trying to measure the costs of these changes, I'd use short
statements executed inside loops in a plpgsql function. And I'd
definitely do it on a machine with quite a few cores (and a running
session on each one), so that spinlock contention might conceivably
happen often enough to be measurable.
If that still says that the runtime cost is down in the noise, we
could then proceed to debate whether min/max are worth a 10% increase
in the size of the shared pgss hash table. Because by my count,
that's about what two more doubles would be, now that we removed the
query texts from the hash table --- a table entry is currently 21
doublewords (at least on 64-bit machines) and we're debating increasing
it to 22 or 24.
regards, tom lane
