I also run my primecoin daemons with the datadir in ram memory to reduce orphans. Every millisecond counts to get a solved block out to the network, don't want any messing about with slow hard drives.
I'm getting a lot of orphans (~1 per hour) so I've been comparing debug.log with blockchain records based on UTC time... In most cases I'm ~1 second late so peer lag is to blame. But a few cases have been within the second, so you're absolutely right that efficiency getting solved blocks out is key.
When running p2pool it is recommended that you don't have too many bitcoind connections.
There's a compromise here: the more connections, the faster your node will be notified of new blocks and avoid wasting work, the faster it can include transactions with fees in the coinbase and the faster it will propagate a P2Pool block minimizing chances it would become orphan. Orphans will quickly rise if you have very few connections (they are the means to be notified of other shares after all).
But the less connections, the less bandwidth used and the lower the latency.
I know we aren't using p2pool but maybe this applies here as well. Too many connections increases chances of your found block getting propagated, but also increases latency. So you need to find a balance.
I've reduced my primecoind connections to 15 with maxconnections=15 command in primecoin.conf.
By the way, I have no proof this works or not. I've got no orphans though.