Institution:
ecoHMEM instruments an application once, profiles its dynamic objects, and generates a tier-aware allocation plant that Flexmalloc enforces at runtime. Developers keep their source code untouched while the framework decides whether each array or buffer should live in fast DRAM/HBM or slower, cheaper persistent memory. This unlocks the extra capacity promised by CXL-ready, pooled-memory nodes and aligns with energy-optimisation goals because off-hot-path data moves to lower-power tiers. Published results show >2x speed-ups on mini-apps and 6% on OpenFOAM over vanilla hardware/software placement, demonstrating tangible benefits for data-heavy HPC and emerging AI pipelines.