On Mon, Jun 29, 2020 at 11:38 AM Fujii Masao
<masao.fujii@oss.nttdata.com> wrote:
>
> >> Your benchmark result seems to suggest that the cause of the problem is
> >> the contention of per-query spinlock in pgss_store(). Right?
> >> This lock contention is likely to happen when multiple sessions run
> >> the same queries.
> >>
> >> One idea to reduce that lock contention is to separate per-query spinlock
> >> into two; one is for planning, and the other is for execution. pgss_store()
> >> determines which lock to use based on the given "kind" argument.
> >> To make this idea work, also every pgss counters like shared_blks_hit
> >> need to be separated into two, i.e., for planning and execution.
> >
> > This can probably remove some overhead, but won't it eventually hit
> > the same issue when multiple connections try to plan the same query,
> > given the number of different queries and very low execution runtime?
>
> Yes. But maybe we can expect that the idea would improve
> the performance to the near same level as v12?
A POC patch should be easy to do and see how much it solves this
problem. However I'm not able to reproduce the issue, and IMHO unless
we specifically want to be able to distinguish planner-time counters
from execution-time counters, I'd prefer to disable track_planning by
default than going this way, so that users with a sane usage won't
have to suffer from a memory increase.
> > I'm wondering if we could instead use atomics to store the counters.
> > The only downside is that we won't guarantee per-row consistency
> > anymore, which may be problematic.
>
> Yeah, we can consider more improvements against this issue.
> But I'm afraid these (maybe including my idea) basically should
> be items for v14...
Yes, that's clearly not something I'd vote to push in v13 at this point.