Extreme-Scale LP Instances in Energy System Analysis: A Benchmark for Shared Memory, Distributed-Memory and GPU accelerated Solvers

This dataset contains a diverse collection of large-scale Linear Programming (LP) instances arising from real-world energy system optimization models (ESOMs). These instances represent the computational frontier in energy research, featuring high spatial and temporal resolutions that challenge current solver technologies and high-performance computing (HPC) environments. These models are derived from various established frameworks and serve as a rigorous testbed for three major solution paradigms: Shared-memory IPMs: Standard industry solvers. Distributed-memory IPMs: Exploiting block-angular structures on HPC clusters GPU-accelerated First-Order Methods (FOMs): Utilizing massive parallelism for ultra-large systems. Following the principles of transparent benchmarking, the test suite is divided into: New Instances: Previously unpublished models consolidated for this repository. Reference Instances: Established baseline problems among others from the Open Energy Transition (OET) benchmark. Note: Confidential industrial instances mentioned in the accompanying paper are excluded from this public release due to proprietary regulations. Problem Class: Pure Linear Programming (LP) formulations. Problem Types: Includes operational dispatch models (inter-temporal coupling via storage) and investment-oriented capacity expansion models (global coupling via capacity variables). Complexity: The number of Non-Zeros in the constraint matrices range from 30 million to 2.68 billion non-zeros. Up to 215 million constraints (rows) and 267 million variables (columns).