iCR: Intent-Driven Adaptable Checkpoint/Restart

Abstract

Modern High-Performance Computing workflows rely on Checkpoint/Restart (C/R) mechanisms for execution continuity, but their inherent task and data heterogeneity demands adaptability from the underlying C/R solution. In a case study with Montage, C/R with static configurations leads to an I/O overhead of $1.6 \times$ and a storage footprint of 176 GB, compared to C/R with an adapted configuration. To this end, we develop intent-driven C/R (iCR), a toolkit that leverages I/O intents as a semantic abstraction that bridges high-level workflow behaviors with low-level C/R policies. Through a trace-driven pipeline, iCR extracts I/O intents from workflows and maps them to representative C/R configurations using a scoring model. This way, iCR reduces the complexity of manual adaptation by 7× while systematizing fine-grained, per-task checkpointing policies that are seamlessly deployed atop popular C/R libraries. Through experiments, we show that seven reusable intents can be extracted from six representative workflows and enable workflow-agnostic adaptable checkpointing policies. Our evaluation demonstrates that, with I/O intents, iCR is able to reduce I/O overhead by up to $2.6 \times$ and improves storage efficiency by $1.8 \times$xb compared to static C/R configurations for heterogeneous workflows such as Montage.