Thanks Tom, that makes a lot of sense. Given we're seeing low iowait and blk_read_time at 4GB shared_buffers, sounds like we should just declare victory here and be happy with that setting?
On Wed, Dec 14, 2022 at 10:27 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Jordan Hurwich <jhurwich@pulsasensors.com> writes: > I'm familiar with the article you linked to, and part of my surprise is > that with these 32GB RAM machines we're seeing better performance at 12.5% > (4GB) than the commonly recommended 25% (8GB) of system memory for > shared_buffers. Your notes about disk read stats from Postgres potentially > actually representing blocks read from the OS cache make sense, I just > imagined that Postgres would be better at managing the memory when it was > dedicated to it via shared_buffers than the OS (obviously with some point > of diminishing returns); and I'm still hoping there's some Postgres > configuration change we can make that enables better performance through > improved utilization of shared_buffers at the commonly recommended 25% of > system memory.
Keep in mind that 25% was never some kind of golden number. It is a rough rule of thumb that was invented for far smaller machines than what you're talking about here.