Adaptive Data Migration in Load-Imbalanced HPC Applications

Distributed parallel applications need to maximize and maintain computer resource utilization and be portable across different machines. Balanced execution of some applications requires more effort than others because their data distribution changes over time. Data re-distribution at runtime requires elaborate schemes that are expensive and may benefit particular applications. This dissertation discusses a solution for HPX applications to monitor application execution with APEX and use AGAS migration to adaptively redistribute data and load balance applications at runtime to improve application performance and scaling behavior. This dissertation provides evidence for the practicality of using the Active Global Address Space as is proposed by the ParalleX model and implemented in HPX. It does so by using migration for the transparent moving of objects at runtime and using the Autonomic Performance Environment for eXascale library with experiments that run on homogeneous and heterogeneous machines at Louisiana State University, CSCS Swiss National Supercomputing Centre, and National Energy Research Scientific Computing Center