The problem is actually that the clients don't really redownload the scratchpad.bin via http but rather they request more and more blocks via json-rpc, via the miner connection from the pool with every round. At some point this request is so big that is bogs down the pool connection because every miner is requesting > 100 blocks via RPC rather than via http...

This is also the problem which other pools considered as "DDoS"... It's no DDoS, it's just a sloppy written miner

On Suprnova you can request max 50 old blocks via RPC, if you request more your connection gets dropped and you have to redownload the scratchpad.bin which is then good for the next 50 Blocks if the miner program doesn't handle the updates well.