Work

One User Dominates: An Audit and User-Grouped Re-check of Task-Failure Prediction on the Alibaba PAI GPU Trace (cluster-trace-gpu-v2020)

Year: 2026
Type: preprint
Abstract: The Alibaba PAI 2020 GPU trace (cluster-trace-gpu-v2020) is used to train and evaluate task-failure predictors. We audit its seven raw tables and quantify two properties that shape any such evaluation. First, one user submits 28.71% of the 985,358 resolved GPU tasks and 49.21% of their failures. Second, some columns nearly copy the outcome, as the KANVAS project's code already notes: the task end time is missing for 99.74% of Failed tasks and no Terminated task, and a task's own machine-metric rows exist for 0.53% of Failed tasks against 76.16% of Terminated ones. We then predict failure from the resource request alone under three split protocols. On random splits gradient boosting has a mean AUC of 0.7950. With whole users held out, the means are 0.5452 for gradient boosting and 0.5226 for logistic regression. KANVAS already reported this collapse on an 18,890-task subset (about 0.45); we extend it to the full trace. On a time-ordered split, logistic regression scores 0.8213 and gradient boosting 0.2811, traced to one request shape of the largest user; with that user removed, gradient boosting scores 0.6714. We also re-check two results from Paper 1 (DOI 10.5281/zenodo.24294903): the 8-GPU vs 2-GPU failure-rate gap holds on the full trace; the quarter-GPU result needs qualifying (47.36% full trace, 15.28% without the largest user). Code, aggregate results and the paigpu module are on GitHub; no trace rows are released.
Language: English
Open Access status: green
OpenAlex ID: W7235362502
Cites: 1
Cited by:
Related to: 10