Workshop: 15th International Workshop on Runtime and Operating Systems for Supercomputers (ROSS 2026)
Authors: Clarete Riana Crasta and Samantika Sury (HPE)
Abstract: Resilience has become significantly more complex for AI workloads. AI models distribute parameters, communication state, and inference caches across thousands of devices. Collective communication has moved onto the critical path of application execution, with large-scale training synchronizing at every step and parallel inference invoking collectives very often. In these environments, a single device or rank failure can affect training progress, model quality, and inference latency. Also, large-scale training and latency-sensitive inference now run on mixed CPU, GPU, and accelerator systems with multiple fabrics and tiers of storage. Large-scale scale-up, scale-out and heterogeneous systems all exacerbate the problem by spreading state across tightly coupled nodes and distributed domains.
Checkpoint/restart remains essential, but it is no longer sufficient as a stand-alone strategy for AI workloads. This paper treats resilience as an optimization problem and proposes a cross-stack resilience architecture that coordinates recovery across applications, framework, communication, storage, orchestration and system telemetry.
Back to 15th International Workshop on Runtime and Operating Systems for Supercomputers (ROSS 2026) Archive Listing
Back to Full Workshop Archive Listing