Towards Dynamic Resource Management with Charm++

Dynamic Resource Management (DRM) increases scheduling flexibility, improving both system performance and individual job execution.

Supporting DRM, however, requires applications to adapt to dynamic changes in allocated resources. Higher-level programming models, such as task-based programming, can simplify this by hiding the complexity of process management and data redistribution.

In this work, we integrate the adaptive features of the Charm++ programming model into the generic Dynamic Processes with PSets (DPP) design as a step towards mainstreaming the use of DRM on HPC systems. To this end, we implement DynCharm, a PMIx-based proxy process manager between Charm++ and the dynamic resource manager, and demonstrate its applicability in DRM scenarios on up to 16 nodes.