"Tom Lane" <tgl@sss.pgh.pa.us> writes:
> I seem to recall that we've previously discussed the idea of letting the
> clock sweep decrement the usage_count before testing for 0, so that a
> buffer could be reused on the first sweep after it was initially used,
> but that we rejected it as being a bad idea. But at least with large
> shared_buffers it doesn't sound like such a bad idea.
I seem to recall the classic clock sweep algorithm was to just use a single
bit. Either a buffer was touched recently or it wasn't.
I also vaguely recall a refinement involving keeping a bitmask and shifting it
right each time the clock hand comes around. So you have a precise history of
which recent clock sweeps the buffer was used and which it wasn't.
I think the coarseness of not caring how heavily it was used is a key part of
the algorithm. By not caring if it was lightly or heavily used during the
clock sweep, just that it was used at least once it avoids making sticking
with incorrect deductions about things like sequential scans even if multiple
sequential scans pass by. As soon as they stop seeing the buffer it
immediately reacts and discards the buffer.
I would check out my OS book from school but it's on another continent :(
-- Gregory Stark EnterpriseDB http://www.enterprisedb.com