The design of the current desktop/server operating systems is premised on the use of slow magnetic disks. Two recent developments, (i) RAM capacity nearing 264 bytes, and (ii) the introduction of non-volatile memory (NVRAM), provide an opportunity for a complete re-design of traditional Unix-like operating systems. We discuss some of the issues which support that proposition and offer a few suggestions for areas that may benefit from looking back at pioneering work. We then propose a segmented memory model for the MIPS processor.

Virtual Memory was devised in a time of scarce resources. In the coming decade we expect to see physical memory systems populated with 264 bytes of RAM, a fraction of which may be non-volatile. Demand paging is inefficient in such large memories because space (Page Tables) and time (Page Table walks) overheads are too high. We collected execution traces from six applications and characterized their virtual memory behavior with respect to miss rates in references to Translation Buffers (TLBs) and Segment Buffers (SBs). Our measurements indicate that the miss rates for SBs are 2-3 orders of magnitude smaller than for TLBs. In light of these results, we discuss some of the design implications of segmented systems and of SBs.