I'm sorry I have to come back at this, but the solution the list gave helped,
but didn't fully solved my problems...
To sum up:
I have a simple table that has indexes on 2 integer columns.
Data is inserted very often (no updates, no deletes, just inserts): at
least 4000/5000 rows per second. The input for the 2 indexed columns
is very random.
I was experiencing bad insertion times past 20M rows.
Some of you suggested that it might have to do with the indexes being
too big to fit in shared_buffers, so I raised it and, in fact, that solved the
issue... but for another 50M rows: basically at roughly 70M rows I'm
back at poor insertion times (disk always 100% busy).
The indexes at this point are 1.7GB each. My shared buffers is 10GB
(machine has 32GB). So I expect the whole indexes to fit in ram, yet
the disk array is 100% busy...
What's going on? What can I try to get back at good insertion
performance?