Topology-Aware Parallelism for NUMA Copying Collectors

NUMA-aware parallel algorithms in runtime systems attempt to improve locality by allocating memory from local NUMA nodes. Researchers have suggested that the garbage collector should profile memory access patterns or use object locality heuristics to determine the target NUMA node before moving an object. However, these solutions are costly when applied to every live object in the reference graph. Our earlier research suggests that connected objects represented by the rooted subgraphs provide abundant locality and they are appropriate for NUMA architecture. In this paper, we utilize the intrinsic locality of rooted sub-graphs to improve parallel copying collector performance. Our new topology-aware parallel copying collector preserves rooted sub-graph integrity by moving the connected objects as a unit to the target NUMA node. In addition, it distributes and assigns the copying tasks to appropriate (i.e. NUMA node local) GC threads. For load balancing, our solution enforces locality on the work-stealing mechanism by stealing from local NUMA nodes only. We evaluated our approach on SPECjbb2013, DaCapo 9.12 and Neo4j. Results show an improvement in GC performance by up to 2.5x speedup and 37% better application performance.