Dynamic page migration in multiprocessors with distributed global memory

Abstract

Multiprocessor architectures using point-to point interconnects and executing parallel algorithms require careful partitioning and allocation of data and processes. For a given interconnection, optimal allocation may be difficult to achieve in some cases because the data-dependent behavior of some algorithms may be impossible to predict. A mechanism called the pivot mechanism is introduced and described that controls the dynamic migration of data pages between neighboring memory modules during program execution. Flexible data migration serves the dual purpose of making algorithms less machine-specific and making possible efficient algorithm execution that is impossible to achieve by using static data allocation.