On 13/08/10 08:38, Tatsuo Ishii wrote:
>> It's slower than smaller numbers, and if you actually dirty a
>> significant portion of it you can have a checkpoint that takes hours to
>> sync, completely trashing system responsiveness for a good portion of it.
>
> So how much is the reasonal upper limit of shared_buffers at this
> point? If it's obvious, should we disable or warn to use more than
> that number?
Trouble is, there won't be a "reasonable upper limit" ... because it
depends so much on the ratio of memory to I/O throughput, the system's
writeback aggressiveness, etc etc etc.
Personally I've had two Pg machines where one seems to suffer with
shared_buffers > 250MB out of 4GB and the other, which has 8GB of RAM,
wants shared_buffers to be around 4GB! The main difference: disk subsystems.
--
Craig Ringer
Tech-related writing: http://soapyfrogs.blogspot.com/