Towards an Adaptive Runtime System for Cloud-Native HPC

A team from the University of Illinois' Charm++ project has published a paper outlining a new adaptive runtime system designed to solve a fundamental mismatch in computing. High-Performance Computing (HPC) applications, like those for climate modeling or molecular dynamics, are built for static, homogeneous supercomputers. The cloud, however, is dynamic, heterogeneous, and offers cost-saving but volatile resources like spot instances. Traditional programming models like MPI cannot efficiently leverage these cloud advantages, often suffering from performance degradation due to network variability and processor differences.

The researchers demonstrate that the asynchronous, message-driven paradigm of the Charm++ runtime system is uniquely suited to bridge this gap. They present two key contributions integrated into a robust framework. First, they implement rate-aware load balancing within Charm++, which dynamically adjusts work distribution across a mix of CPU and GPU instances based on their real-time performance, mitigating issues like network contention. Second, they extend an existing resource management framework to support GPU and CPU spot instances with minimal interruption overhead when these low-cost instances are reclaimed by the cloud provider.

Together, these tools allow tightly coupled scientific applications to run resiliently on dynamic cloud infrastructure. The system enables applications to automatically adapt to performance variability and heterogeneous processors, while also tapping into the significant cost savings of spot markets. This work provides a practical pathway for migrating and modernizing legacy HPC workloads, making supercomputing-grade performance more accessible and affordable via commercial clouds.