Set up eloipool with the difficulty set to some high value (eg. 64 with ShareTarget = 0x0000000003ffffffffffffffffffffffffffffffffffffffffffffffffffffff). Mine with getwork, not getblocktemplate.
Dynamic targetting was used, which should in theory achieve same... maybe high+static is better.
And make sure roll-n-time is functional. That setup should be able to handle hundreds of Ghash/s easily. That's what I use for my FPGA farm.
IIRC your pushpool and p2pool both hard-code the diff to 1. Isn't it why they can't keep up with the many getwork calls?
Multiple answers...
pushpool has several known inefficiencies... besides being single-threaded for database + network + rpc, it simply proxies every client getwork to bitcoind, with some rewriting. bitcoind cannot keep up with that, once the miner load (historically difficulty 1.0) reaches a certain point, the whole system cannot keep up.
Pool ops either solve this by scaling up the number of bitcoind processes (machines), or run pool server software that supports some amount of internal work generation, easing bitcoind's load.
nothing to be proud of imo.
