ABSTRACT:
Garbage collection is an important features of modern programming
languages: by liberating the programmer from the responsibility of freeing
up unused memory by hand, it leads to code with fewer bugs and cleaner
design. However, these software engineering benefits have their costs:
garbage collection may incur annoying pauses, slowdowns, and increased
memory requirements.
                                                                                
This talk presents a novel family of garbage collection algorithms
that use information about the connectivity of heap objects to reduce
these costs. I first present empirical data showing that connectivity
information is a good indicator for when objects die. Then I describe
a family of garbage collectors that exploit connectivity properties
to yield short pause times, good throughput, and low memory
footprint. As part of implementing this collector, I developed the
first non-trivial pointer analysis that handles all of Java,
including dynamic class loading, reflection, and native code. I
present an evaluation of connectivity-based garbage collection and
the pointer analysis.