Speaker
Description
The offloading of Kubernetes‑managed computing workloads to the serverless INFN DataCloud infrastructure has become a key component in the emerging national strategy for federating heterogeneous computing resources. In this contribution, we present a workflow designed to simulate the response of 3D diamond detectors to minimum‑ionizing particles, providing the ground-truth for validating Physics Informed Neural Networks. This workflow is representative of a broader class of scientific applications because it integrates GPU‑accelerated computation with legacy high‑throughput (HTC) software originally developed for traditional grid environments.
To enable portable execution across heterogeneous backends, we decomposed the workflow into a Directed Acyclic Graph (DAG) managed by Snakemake and deployed through the AI_INFN Platform. This design separates HTC tasks from GPU‑intensive kernels, allowing their distribution across both the INFN Tier‑1 facility and the TeRABIT HPC Bubble in Padova via the interLink federation layer.
We report on the practical limitations encountered when using Snakemake as the orchestrator in this offloading scenario, particularly regarding secret management and automation. Finally, we discuss how recent developments, most notably the introduction of an overlay network in interLink deployments, enable the adoption of modern workflow engines such as Argo Workflows, opening the path to more flexible, cloud‑native orchestration models.