[Veserv]

The answer is memory fragmentation. Process memory safety can theoretically be done with a different scheme such as hardware memory capabilities. You could probably even do demand paging if the memory capability is not a true pointer, but some sort of memory system coherent handle. But, as soon as a process wants to allocate a 1 GB chunk and you have 2 GB, but only in 64 KB chunks you have a problem. You could go around copying data to compact the physical memory, but now you have serious interference problems.

You then run into the next problem of using say 32 MB in a "hot loop" in the middle of a 64 GB demand-loaded data structure on a 32 GB machine. You can not greedily load in the entire data structure from disk, so you need some sort of subset feature on your memory handles. But then what do you do about using two disjoint sections separated by over 32 GB? You need some way of having multiple subsets that correspond to physical addresses that do not respect the handle offset. Subset 1 corresponds to physical address range A and subset B corresponds to a uncorrelated physical address range B. Congratulations, you have reinvented memory mapping with extra steps.