I think there was mis-understanding as I have not discussed the load
balancing to the depth as the original poster just wanted to know from
the tnsnames.ora level. That is why I had clarified him saying TAF is a
feature from the network layer, not from the services layer and TNS
load balancing is purely round robin. It does not care about CPU load
or number of connections.
To answer your question, you are commenting about listener load
balancing where we register the CPU load to the other nodes (using
remote listener) and do the load balancing based on the least loaded
note. And also there is a way to disable that using an underscore
parametetr (_prefer_least_loaded_node or something like that in
listener.ora) as the listener load balancing does not perform
effectively for benchmarks or logon storms.
If you have the listener load balancing setup, the listeners cross
register with other participating instances and update the cpu load in
every 5 minutes. If you have all connections coming up with in 5
minutes there are chances that all connections will go to the first
identified node. I think that is what you have experiened and that is
not based on numbers (10+ connections as you mention), but based on the
CPU load during that time.
There are couple of other load balancing options also available based
on other network level parameters. But to answer your question,
listener load balancing is based on machine load and client load
balancing is based on round robin.
Hope this clarifies this
Co-Author: Oracle Wait Interface, Oracle Press 2004