> | ARC still helps, since it makes sure the shared_buffers don't all get
> | flushed from the useful small datasets when a seq scan gets executed.
>
> I'm still not convinced. Why the last backend alive, have to throw away
> bunch of memory copied in the SHM? And again, the ARC is a replacement
> policy for a cache, which one ?
As you know, ARC is a recent addition. I've not seen any benchmarks
demonstrating that the optimal SHARED_BUFFERS setting is different today
than it was in the past.
We know it's changed, but the old buffer strategy had an equally hard
time with a small buffer as it did a large one. Does that mean the
middle of the curve is still at 15k buffers but the extremes are handled
better? Or something completely different?
Please feel free to benchmark 7.5 (OSDL folks should be able to help us
as well) and report back.