"Ed L." <pgsql@bluepolka.net> writes:
> So, pursuing theory (1) in 'the time before ARC', assuming you have a
> dedicated box with little or no non-db competition for resources, why give
> PG anymore than is absolutely required (max_connections*2 or 16)? Why not
> just let the kernel reign as completely as possible? Is that what you mean
> by "relatively small"?
Well, if you go *really* small then you find a lot of CPU time gets
wasted shuffling data from kernel cache to PG cache. The sweet spot
for theory (1) seems to be to set shared_buffers in the range of 1000 to
10000 buffers. (Time was that that was a serious amount of RAM, but
not any more...)
regards, tom lane