We presented a discussion on address translation strategies, both in the context of the Texas persistent store and for general persistence implementations. We also proposed a new classification for persistence in terms of granularity choices for fundamental design issues rather than using taxonomies based only on address translation semantics, and discussed each choice that we made in Texas.
We also discussed issues related to fine-grained address translation, including their inherent costs that make them unsuitable as the primary address translation mechanism in a persistence implementation. Instead, we discussed how a mixed-granularity approach can be used to selectively incorporate fine-grained address translation in the application.
We presented our implementation of mixed-granularity address translation in Texas which combines the C++ smart pointer idiom for the fine-grained translation component with the normal pointer swizzling at page fault time mechanism for the coarse-grained translation component, while maintaining portability and compatibility of the system.
Our basic performance results using the OO1 benchmark have shown that the mixed-granularity approach works well for applications with data structures that do not provide the best performance with a pure coarse-grained approach. However, further performance measurements are necessary, especially using real applications instead of synthetic benchmarks which do not always model reality very well.