Memory management for parallel tasks in shared memory

Abstract

Three memory management strategics for sl arcd-memory multiprocessors that support two-space copying garbage collection are presented. These strategies exploit the fork-join task structure of the divide-and-conquer paradigm by equipping each task with a private heap that can be locally collected independently of other processors and tasks. The memory management strategies use a virtual address space to allocate private heaps such that the efficient copying collectors do not need to be adapted to handle physically scattered heaps. When the allocation strategies run out of the virtual address space, an expensive compaction operation has to be performed. Results from a detailed simulation, however, show that this happens so infrequently that the costs are negligible in practice.