A study of sharing definitions in thread-local heaps (Position Paper)

With the advent of larger heaps, multi-core processors and NUMA architectures, garbage collection scalability is evermore important. Shared memory is an important bottleneck and stress on shared memory can be reduced by using Thread-local heaps. Thread-local heaps provide a promising solution to this challenge, distinguishing between local objects that do not escape their allocating thread and shared objects that do. This allows a new type of collection that requires a single thread’s co-operation and affords more intelligent object placement in the heap. We examine options for their design and suggest a new design.