On Wed, May 7, 2014 at 11:04 AM, Josh Berkus <josh@agliodbs.com> wrote:
On 05/06/2014 10:35 PM, Peter Geoghegan wrote: > +1. In my view, we probably should have set it to a much higher > absolute default value. The main problem with setting it to any > multiple of shared_buffers that I can see is that shared_buffers is a > very poor proxy for what effective_cache_size is supposed to > represent. In general, the folk wisdom around sizing shared_buffers > has past its sell-by date.
Unfortunately nobody has the time/resources to do the kind of testing required for a new recommendation for shared_buffers.
I think it is worse than that. I don't think we know what such testing would even look like. SSD? BBU? max_connections=20000 with 256 cores? pgbench -N? capture and replay of Amazon's workload?
If we could spell out/agree upon what kind of testing we would find convincing, that would probably go a long way to getting some people to work on carrying out the tests. Unless the conclusion was "please have 3TB or RAM and a 50 disk RAID", then there might be few takers.