I have a game running on GAE (www.runesketch.com), our initial design was really ineffecient. Using Khan Academy's mini profiler (https://github.com/kamens/gae_mini_profiler) we slashed our gets, puts and general overheads by two orders of magnitude. Our main cost, running a live game, we managed to completely remove gets and puts and run in memcache. We hardly scratch our daily usage now.

One of the overheads I did not realise that took a long with was using channels. We found them about 3x slower than a get or put. By parrallelising those with another thread we reduced our latency hugely. Our game went from taking 8 seconds to service a move, to 300ms.

Just my experience, did you work hard with a profiler before moving all the code?