> I don't think this will be material; the improvement is primarily in going
> from 1 to 100; the improvement from going to (say) 1000 is small, in
> comparison.  And I think 1.2 improved this in the logic that deals with large
> tuples...
problem for me are not big tuples but lot of small ones. tracing network
discovered that fetching 100 rows needs just 11 1500 sized network packets.
Because my network is fast > 100Mbit,  but it has high RTT time ~ 270 ms
bandwith is not a problem and packet loss is < 0.1%. Switching from cursor
fetching to COPY will provide major speedup.

> The "better" answer would involve trying to do "peephole optimization"
> on the query stream so as to group together related updates. 
yes, all high-end replicators for Oracle are using it.