Re: [Patch][WiP] Tweaked LRU for shared buffers

On 2/16/19 12:51 AM, Peter Geoghegan wrote:
> On Fri, Feb 15, 2019 at 3:30 PM Tomas Vondra
> <tomas(dot)vondra(at)2ndquadrant(dot)com> wrote:
>> That TPS chart looks a bit ... wild. How come the master jumps so much
>> up and down? That's a bit suspicious, IMHO.
>
> Somebody should write a patch to make buffer eviction completely
> random, without aiming to get it committed. That isn't as bad of a
> strategy as it sounds, and it would help with assessing improvements
> in this area.
>
> We know that the cache replacement algorithm behaves randomly when
> there is extreme contention, while also slowing everything down due
> to maintaining the clock.

Possibly, although I still find it strange that the throughput first
grows, then at shared_buffers 1GB it drops, and then at 3GB it starts
growing again. Considering this is on 200GB data set, I doubt the
pressure/contention is much different with 1GB and 3GB, but maybe it is.

> A unambiguously better caching algorithm would at a minimum be able
> to beat our "cheap random replacement" prototype as well as the
> existing clocksweep algorithm in most or all cases. That seems like a
> reasonably good starting point, at least.
>

Yes, comparison to cheap random replacement would be an interesting
experiment.

regards

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