A study of connected object locality in NUMA heaps

Reference locality is vital to the performance of parallel Garbage Collection (GC) running on Non-Uniform Memory Access (NUMA) machines. A GC thread may trace remotely placed objects that descend from the root set or, for load balance, a GC thread may steal non-local objects from other threads' work lists. Processing distant live objects could introduce contention in the interconnect links between memory nodes and it could increase memory access latency. Researchers have proposed various techniques to improve GC tracing locality. However, few studies attempt to optimize the locality of connected objects in NUMA object graph. In this paper, we study the locality of a rooted subgraph, a unit of object connectivity in the object graph. A rooted subgraph is a set of references containing one root reference, and every reference is reachable from the root. We empirically study the locality of rooted subgraphs of DaCapo and SPECjbb2005 benchmark suites. The results show that more than 80% of objects in a rooted subgraph are located in the same memory node as the root object. We then propose a GC locality optimization that uses the root memory node as a heuristic to guide GC threads processing local objects.