So let me ponder a bit on my favourite (rumoured) NX aspect, slow monday and all.
SCD, aka supplements.. That thing can fly. Especially when configured in a local network with the main unit. I've been experimenting lately with such setups for a pet project of mine, using highly asymmetrical protocols (and bypassing TCP/IP altogether but that's another story) where the supplement has all assets in advance (in my case - procedural assets), and the main unit sends out a single beacon/sync message when ready to accept during a frame, while in the meantime the supplement has finished or is about to finish its work, and at said beacon just sends away all done computations for the frame, and then immediately starts with the next frame. Given a proper heuristics for the work split between main and supplement, the two nodes achieve perfect utilisation. And just to give you a clue how asymmetrical it all is (in packets) - it's kilobytes from the main-unit side vs gigabytes from the supplement side (over a GbE link). In my case the workload has seen perfect linear scaling thanks to the precise workload split between the nodes.
So, two things make for a great distributed system:
- running a minimal overhead sync protocol over a low-latency network connection
- optimal workload split between the nodes, so no idle time on either end.
Satisfying those yields perfect (read: linear) gains, aka profit : )