Title: Optimizing thread-local variables in Java 

Authors: Yi Zhang, Clark Verbrugge 
Speaker: Yi Zhang

Abstract:

Many of the new concurrent programming languages, such as OpenMP, UPC, and 
D, include features that make heavy use of thread-local data. When porting 
such features as part of a language extension to Java, however, performance 
can suffer due to the simple, map-based implementation of the ThreadLocal
class.  We implement an optimized mechanism based on programmer annotations 
that can efficiently ensure class and instance variables are only accessed 
by their owner thread. Both class and instance variables inherit values from 
the parent thread through deep copying, allowing all the reachable objects 
of child threads to have local copies if syntactically specified. In 
particular, class variable access involves direct access to thread-local 
variables through a localized heap, which is faster and easier than 
searching a map.  Our design improves performance significantly over the 
traditional thread-local access method for class variables. We further 
evaluate our approach by modifying non-trivial, existing benchmarks to make 
better use of thread-local features, illustrating feasibility and allowing 
us to measure the improvement in realistic contexts.  This work is intended 
to bring us closer to designs for a complete race-free version of Java, as 
well as show improved support for use of thread-local data in other languages.