Guillaume,
> which shows the index on primary keys is used, but is always read
> from disk.
or, more likely, from the FS cache.
> But, the clock time used for the request is actually identical
> when using -B 1000 or -B 20000. I suppose the kernel is bringing
> the performance difference thanks to filesystem caching.
Yes. The only way you'd see a differeence is on a mixed load of
concurrent read & write queries. Any single-query test is unlikely to
show a difference between using the FS cache and shared_buffers.
--Josh