next up previous
Next: Virtual-Processor Model Up: Kernel Support Previous: Kernel Support

Memory Model

A Nemesis kernel provides a number of distinct, schedulable entities, called domains. While all domains share the same virtual address space, privacy and protection are implemented using the appropriate access rights in the virtual address translations. Code executing within a domain may access memory within another domain only if both domains have explicitly arranged to share the memory.

Some examples highlight the approach: shared library segments would be mapped readable in every domain; a unidirectional inter-domain communications channel would be mapped read/write in the source and read-only at the sink; objects may be shared in shared read/write segments; etc.

The cost of using a single address space is the penalty of load-time relocation. We try to amortise this cost by caching the results of such relocations and then aim to reload an application at the same virtual address at which it was last executed. In this we are helped by the use of 64-bit VM architectures, which allow a sparse allocation of addresses so that we can arrange reuse with high probability. Consider for example allocating the top 32 address bits of a 64 bit virtual address based on a 32-bit hash function of the code to be executed.

The benefits of a single address space we are aiming for are: simplified sharing of data structures (in particular objects) between domains, and the removal of virtual address aliases which can result in significant context switch costs with caches accessed by virtual address.

Sape J. Mullender, Ian M. Leslie and Derek McAuley