Persistent programming languages provide object persistence across program invocations, by treating volatile memory as a cache for stable storage. Orthogonal persistence allows any object to be potentially persistent, without any restriction on its type. Adding orthogonal persistence to a language environment presents several performance-related challenges. This work is aimed at reducing the various overheads associated with orthogonal persistence. The costs being targeted are read barriers, write barriers and swizzle barriers. A read barrier checks the cache residency of the target object while a write barrier marks the target as dirty in the cache. Many of these read and write barriers are redundant, and applying partial redundancy elimination of pointer-based access path expressions can be very beneficial in eliminating them. Swizzling is the translation of an object reference from an external, persistent format to an internal, transient format; a swizzle barrier checks and makes sure that a reference is swizzled. Swizzle barriers have the added overhead that they are usually associated with iteration over container objects like arrays. Hence, there is a need for an additional optimization to merge multiple swizzle barriers into one inclusive barrier. By induction variable analysis, the bounds of a variable being used to loop through an array can be determined. We exploit this information to develop a range swizzle optimization technique to reduce the overhead of swizzle barriers. We have implemented our analysis and optimization framework for an orthogonally persistent Java system. In experiments performed on several benchmarks and applications, our optimizations eliminated on average 83 percent of read barriers, 25 percent of write barriers and 66 percent of swizzle barriers.
@mastersthesis{Brahnmath1998MS,
author = {Brahnmath, Kumar Jagadeeshwaraiah},
title = {Optimizing Orthogonal Persistence for Java},
school = {Purdue University},
year = {1998},
month = {May}
}