A Workload-custom System-informed Approach for Leveraging Future Disaggregated Memory Systems

Today´s HPC systems use tightly coupled memory and compute resources within compute nodes and rely on overprovisioning to support diverse workloads. However, recent studies show that such systems could result in significant energy and resource waste. Thus, they are not sustainable due to environmental impact.

An alternative is disaggregated systems with loosely coupled compute, memory, accelerator, and storage resources.

Disaggregated systems can allocate resources at fine-granularity based on workloads´ needs to achieve resource efficiency.

This project targets disaggregated memory, a critical part of next-generation energy-efficient computing infrastructures.

The project aims to enable the efficient execution of industrial and scientific workloads on future computing infrastructures with disaggregated memory.

The primary approach combines workload characteristics and runtime system states to optimize data placement and movement.

The novelty is to capture composable memory requirements in a workload in logical data views and use them to guide memory management and data path selection.

This project will conduct four research tasks: discovering and leveraging composable memory requirements for workload-informed memory optimizations, developing software-controlled transient data caching services and performance-aware along-the-path data reduction, and exploring global coordination based on system states.