MUMBAI, India, Aug. 12 -- Intellectual Property India has published a patent application (202621075941 A) filed by Symbiosis International Deemed University on June 18, 2026, for Leakage-Aware Machine Learning System For Submission-Time Gpu Job Runtime Prediction In Cloud Datacenters.
Inventor includes Dr. Shreyas Rajendra Hole.
The application for the patent was published on August 07, 2026, under issue no. 32/2026.
Abstract: ABSTRACT LEAKAGE-AWARE MACHINE LEARNING SYSTEM FOR SUBMISSION-TIME GPU JOB RUNTIME PREDICTION IN CLOUD DATACENTERS A deployment-aware machine learning system (100) and method for predicting GPU job runtime at submission time in large-scale cloud and HPC datacenters is disclosed. The system comprises a data ingestion and preprocessing module (110) that transforms raw scheduler records and GPU-linked traces into a job- level dataset through deduplication, merge, and aggregation. A leakage-aware feature engineering module (120) enforces a strict temporal boundary excluding post- execution variables and constructs GPU-centric resource-ratio features including CPU-to-GPU and memory-to-GPU ratios. A log-transform target module (130) handles heavy-tailed runtime distributions. A stacked ensemble prediction engine (140) combining LightGBM, XGBoost, and CatBoost achieves R-squared of 0.71 on log-transformed targets. A SHAP-based interpretability module (150) provides feature attribution for model transparency. A scheduler integration interface (160) implements a trust-but-verify policy comparing predicted runtime and confidence intervals against user-declared walltime to support backfilling, reservation sizing, and proactive queue management for improved GPU utilization. [
Disclaimer: Curated by HT Syndication.
