...so it can't be too much harder than PPLNS to implement here - especially when Slush has more experience than anyone at running a pool using exponential scoring).
The main difference between current score implementation is that score equation needs only time as an input, but proper implementation of DGM need to know also count of already submitted shares (across all backends). Pool core is designed quite scalable, there's no online synchronization between backends necessary, but it goes against DGM requirements. And it's the pain and the reason why it takes so much time. Originally I expect that it won't be so hard, but I felt into solving strange threading issues (cause framework which I'm currently using doesn't natively support user's threads).
Can't you multiply (or add the logs) of all the "s" variables from all the backend nodes and then do the same for the per-user scores (sk) ?
It might work, but I suck at math, Meni may have an idea on how to algebrize that.
Edit: Just in case I was not clear, the idea would be to have each backend node do DGM independently and then multiply all the backend "s" together and all the backend "sk" together to estimate or compute the payout.