External memory data structure to replace vector of maps

I'm doing an iterative computation on a flow network during which I need to record how much each source contributes to the flow on each edge. The flow on any one edge is due to 2% of the sources on average, so I define `vector< map<int, double> > flow`, where `flow[e][s] = f` means that the flow on edge `e` due to source `s` is `f`. At each iteration, every `f` in `flow` gets updated. The program's peak memory usage gets close to 4 GB. This works on (32-bit) Linux and OS X, but it crashes on Windows (which seems to impose a 2 GB per process limit). How can I implement a disk-based data structure with a `vector< map<int, double> >` interface (or otherwise get around this problem)?