I'm inclined to believe the memory leak... although I've only seen it on my Linux (Ubuntu 14.04) node. I watched bitcoind spool up to consume 8GB of RAM before it was shut down. On my Mac, using the same codebase, I regularly only see about 400MB of RAM being consumed by the bitcoind. Both Linux and Mac are built from source.
Are both bitcoind nodes being used for p2pool? My testing indicates that the use of getblocktemplate is necessary for the memory leaking to occur.
Yes. I was running a VPS (a DigitalOcean droplet) that I used as my backup p2pool node. I had other miners besides my own on it pretty constantly. It ran for quite a long time until a few weeks ago it started acting up. The bitcoind process would die for no apparent reason. Sometimes the p2pool process would just die off as well. Finally spending time looking at the box, I watched as the bitcoind would slowly consume more and more memory, up until it took almost the entire 8G available, at which point it would invariably die off.
On my Mac, I also run a p2pool node that is local to my miners at my home. That bitcoind, just like the one on my droplet, was compiled from source. On my local node, I've never seen it utilize more than 1G of RAM. More typically, it's around 400-500MB. Granted, I have fewer miners on my local home node (2-3 at most with a total hash of about 5TH/s). On my Linux node, I typically had 5-6 miners with a hash rate of about 15TH/s.