Re: Keep elog(ERROR) and ereport(ERROR) calls in the cold path

On 2020-09-22 22:42, David Rowley wrote:
> On Tue, 22 Sep 2020 at 19:08, David Rowley <dgrowleyml(at)gmail(dot)com> wrote:
>> I ran another scale=5 TPCH benchmark on v4 against f859c2ffa using gcc
>> 9.3. I'm unable to see any gains with this, however, the results were
>> pretty noisy. I only ran pgbench for 60 seconds per query. I'll likely
>> need to run that a bit longer. I'll do that tonight.
>
> I've attached the results of a TPCH scale=5 run master (f859c2ffa) vs
> master + elog_ereport_attribute_cold_v4.patch
>
> It does not look great. The patched version seems to have done about
> 1.17% less work than master did.

I wonder how much benefit you'd get from

a) compiling with -O3 instead of -O2, or
b) compiling with profile-driven optimization

I think that would indicate a target and/or a ceiling of what we should
be expecting from hot/cold/likely/unlikely optimization techniques like
this.

--
Peter Eisentraut http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services