Does nsd still zombie on 2.4 kernels when a handle goes stale? The traditional way to avoid the zombie problem is to set the lifetime of database handles to a very large value so once connected to Oracle they'll never be allowed to die.
For production systems this seems like the best thing to do anyway (I know aol.com runs with db handles configured to live forever and if it works for them it will work for me) so try this if you've not already.
