I've always wondered what the development process looks like for these type of algorithms. If you have to run the program for a year to know if it will work, how can you have any confidence that what you've written is going to do the trick?
It brings interesting trade-offs for program design. You can write the code one way which may be 10x faster but harder to reason about, or another way which is more straightforward but takes an extra 5 days go execute. How confident are you in your code or debugging ability? How many iterations will you need? I'm assuming this was written in CUDA based on the block/thread ID mix-up.