In the case when shared_buffers cover most of RAM, most of writes should happen by checkpointer, and cache hit ratio should be high. So a hypothetical question: Could shared_buffers=200GB on a 250 GB RAM server ever be a reasonable setting? (assuming there are no other applications running except postgres, and 50GB is enough for allocating work_mem/maintenance_work_mem and for serving queries)
That amount of shared buffers is not sensible. I found on a 256 GB box that anything over about 50-100GB was counter productive. That was a FreeBSD system where I ran the database on top of ZFS, so there was a lot of data also stored in the ARC (memory cache). There is a setting in postgres to tell it how much RAM your system is using for the disk cache, so set that to a fair estimate of how much your system will use. I set mine to 50% of RAM. I did not limit the cache at the OS level since it is good about giving up that memory for the needs of the running processes.
