I said:
> This doesn't seem to quite square
> with your explanation though --- surely the number should go to 8000 and
> change? The man page for top says these numbers are in kilobytes ...
> but if they were really measured in, say, 4K pages, then we'd be talking
> about 26M of shared memory touched, which might be plausible when you
> consider shared libraries.
Never mind --- further testing shows that top does report in kilobytes.
I made a silly mistake in writing my test query that prevented it from
using as many buffers as I expected. When I write something that really
does use all 1000 buffers, SHARE goes to 10392, which is right about
what you'd expect.
So I think this mystery is solved. Back to chasing real bugs ...
regards, tom lane