=?ISO-8859-1?Q?I=F1igo?= Martinez Lasala <imartinez(at)vectorsf(dot)com> writes:
> We have increased shared buffers to 2.5GB (linux kernels allows us reach
> this level) and lowered work_mem to 500MB.
You are apparently unclear on the concept. Pushing the database to the
limit of the available address space is probably going to result in
failures. Redistributing a set of overoptimistic parameters into a
different set of overoptimistic parameters will not fix this.
regards, tom lane