I was originally a bit confused about how to determine the yellow regions, but I later realized the "order" of events would probably be like the following: Step 1, figure out which are the current vertices we are looking at (in this case 3,4) so we pull the entire second shard into memory. Then Step 2, we loop over shards 1 and 3 (src row), and we only pull the contiguous memory blocks we needed.

Just to point out something that could be potentially interesting: the optimization via locality could even be done at the register level, i.e., considering what could fit into the registers.