The importance of locality and sparsity in memory management
Our previous articles describing the disadvantages of generational garbage collection and our prototype mark-region memory management system designed for HLVM originally showed that region-based allocation and deallocation has the potential to be only 4-20% slower than OCaml's generational collector. However, our more recent work that was designed to be more realistic by deferring deallocations to bulk GC cycles was significantly slower, around twice as slow as OCaml. There are several differences between the stack-based deallocation scheme used in the first benchmark and the GC-like deallocation scheme used in the second benchmark that have the potential to account for this performance degradation: The mock GC introduced mark bytes into the allocated values and marked them as unreachable when they fell out of scope in the mutator. The mock GC allocated by popping a reference of the top of a region's free list. The mock GC deallocated by pushing a reference onto the free lis...