What an odd coincidence. Just yesterday, I was benchmarking the overhead of turning on the MMU on an embedded system. It was about a 10% hit in performance, a

What an odd coincidence. Just yesterday, I was benchmarking the overhead of turning on the MMU on an embedded system. It was about a 10% hit in performance, and this is without paging or even changing the mapping at all. In this paper, we envision a _physically addressed_software architecture, without address translation, in which the OS hands out fixed-size blocks of memory, and applications allocate data structures across those blocks. This is essentially nearly every embedded system I've ever worked on. The dumbest possible memory allocator starts at a base and increments, never freeing memory except possibly on a major frame.