ADAPt: Exploiting Abstract Data-Access Patterns For Better Data Locality In Parallel Processing
The ADAPt project, funded by the Research Council of Norway, aims to develop new techniques to improve data locality in parallel processing. The idea is to systematically coordinate task scheduling and data allocation using predicted patterns of data access to memory. The approach combines formal models of parallel systems with basic research in programming language theory to (1) capture abstractly the interaction of workflows with dynamically created tasks and memory locations on parallel computers, and (2) combine formal analysis with model-based simulations to explore how to systemically calibrate schedulers and allocators for specific applications.
Contacts: Ferruccio Damiani