On Sat, Oct 8, 2016 at 12:58 PM, Fabien COELHO <coelho@cri.ensmp.fr> wrote:
>
> Hello Tom,
>
> I comment here on the first part of your remarks. I've answered the second
> part in another mail.
>
>>> (1) The required schema is slightly different : currently the type used
>>> for holding balances is not wide enough per the TCP-B standard, this mean
>>> maybe having an option to do "pgbench -i --standard-tpcb" which would
>>> generate the right schema, probably it should just change a few INTEGER
>>> to
>>> INT8, or maybe use NUMERIC(10). I have not done such a patch yet.
>>
>>
>> The whole question of the database setup is an interesting one.
>> If we were to do anything at all here, I'd want to see not only the table
>> schemas and initial population, but also the hard-wired "vacuum" logic,
>> somehow made not so hard-wired. I have no good ideas about that. The SQL
>> commands could possibly be taken from scripts, but what of all the work
>> that's gone into friendly progress reporting for table loading?
>
>
> I'm unconvince by the current status, especially the default behaviors. I
> think it should do a good sensible representative job, and not be a minimum
> installation.
>
> For instance, the default setup does not use foreign keys. It should be the
> reverse, foreign keys should be included by default and an option should be
> used to lessen the schema quality.
>
> Also, given the heavy UPDATE nature of the pgbench test, a non 100% default
> fill factor on some tables would make sense.
>
FWIW, sometime back I have seen that with fill factor 80, at somewhat
moderate client counts (32) on 192 - Hyper Threaded m/c, the
performance is 20~30% better, but at higher client counts, it was same
as 100 fill factor. I think if go by your theory, one could also
argue to have non-default values autovacuum threshold parameters.
pgbench already has a parameter to specify non-default fillfactor and
I think that is sufficient for anyone to do performance testing.
--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com
