On Sat, Oct 11, 2025, at 08:43, Joel Jacobson wrote:
> In addition to previously suggested optimization, there is another major
> one that seems doable, that would mean a great improvement for workload
> having large traffic differences between channels, i.e. some low traffic
> and some high traffic.
>
> I'm not entirely sure this approach is correct though, I've might
> misunderstood the guarantees of the heavyweight lock. My assumption is
> based on that there can only be one backend that is currently running
> the code in PreCommit_Notify after having aquired the heavyweight lock.
> If this is not true, then it doesn't work. What made me worried is the
> exclusive lock we also take inside the same function, I don't see the
> point of it since we're already holding the heavyweight lock, but maybe
> this is just to "allows deadlocks to be detected" like the comment says?
..,
> * 0003-optimize_listen_notify-v14.patch:
>
> Optimize LISTEN/NOTIFY by advancing idle backends directly
>
> Building on the previous channel-specific listener tracking
> optimization, this patch further reduces context switching by detecting
> idle listening backends that don't listen to any of the channels being
> notified and advancing their queue positions directly without waking
> them up.
...
> 0003-optimize_listen_notify-v14.patch:
>
> idle_listeners round_trips_per_sec max_latency_usec
> 0 33236.8 1090
> 10 34681.0 1338
> 20 34530.4 1372
> 30 34061.6 1339
> 40 33084.5 913
> 50 33847.5 955
> 60 33675.8 1239
> 70 28857.4 20443
> 80 33324.9 786
> 90 33612.3 758
> 100 31259.2 7706
I noticed the strange data point at idle_listeners=70.
This made me think about the "wake tail only" trick,
and realized this is now unnecessary with the new 0003 idea.
New version attached that removes that part from the 0003 patch.
This also of course improved the stability of max_latency_usec,
since in this specific benchmark all other listeners are always idle,
so they don't need to be woken up ever:
idle_listeners round_trips_per_sec max_latency_usec
0 33631.8 546
10 34318.0 586
20 34813.0 596
30 35073.4 574
40 34646.1 569
50 33755.5 634
60 33991.6 561
70 34049.0 550
80 33886.0 541
90 33545.0 540
100 33163.1 660
/Joel
