Coupling Memory and Computation for Locality Management

摘要: We articulate the need for managing (data) locality automatically rather than leaving it to programmer, especially in parallel programming systems. To this end, we propose techniques coupling tightly computation (including thread scheduler) and memory manager so that data can be positioned closely hardware. Such tight of management is sharp contrast with prevailing practice considering each isolation. For example, memory-management usually abstract as an unknown "mutator", which treated a "black box". As example approach, paper consider specific class computations, nested-parallel computations. computations dynamically create nesting tasks. method organizing tree heaps reflecting structure nesting. More specifically, our approach creates heap task if separately scheduled on processor. This allows us couple garbage collection way processors. enables taking advantage program by mapping improved collected immediately after its finishes when contents likely cache.