Curt Sampson <cjs@cynic.net> writes:
> On Fri, 28 Jun 2002, Tom Lane wrote:
>> And the swapping activity is exactly the problem, isn't it?
> That particular swapping activity would not be a problem. Memory
> that's not used gets paged out and that's the end of it. The problem
> is that something is *using* that memory, so it's not being paged
> out, or if it does get paged out, it gets paged back in again.
Yeah --- but typical implementations of malloc are very pager-
unfriendly; they tend to traverse data structures that consist
of a word or two at the head of each randomly-sized chunk of
data or former-now-freed data.
PG adds its own layer of not-very-paging-friendly allocation
logic on top of whatever sins your local malloc may commit.
Bottom line is that a PG backend that's swollen to a couple
hundred MB is trouble. Don't assume it'll play nice with the
swapper; it won't.
regards, tom lane