Josh Berkus <josh@agliodbs.com> writes:
>> If we had a portable way
>> of preventing the kernel from caching the same page, it would make more
>> sense to run with large shared_buffers.
> Really? I thought we wanted to move the other way ... that is, if we could
> get over the portability issues, eliminate shared_buffers entirely and rely
> completely on the OS cache.
That seems unlikely to happen: there are cache-coherency problems if you
don't do your page-level access through shared buffers. Some have
suggested using mmap access to the data files in place of shared memory,
but that introduces a slew of issues of its own. It might happen but
I'm not holding my breath.
regards, tom lane