ATLAS: Alibaba Dataset and Benchmark for Learning-Augmented Scheduling

The paper introduces ATLAS, a production-derived dataset for learning-augmented scheduling, along with a prediction benchmark and a multi-stage ML model. It resides in the 'Benchmarks, Datasets, and Evaluation Frameworks' leaf of the taxonomy, which contains only two papers total. This leaf is notably sparse compared to more crowded branches such as 'Cloud and Data Center Scheduling' (six papers) or 'Task Execution Time and Resource Prediction' (four papers). The scarcity of benchmark resources in this field underscores the potential value of a well-curated dataset, as most prior work has focused on algorithm design or prediction models rather than standardized evaluation infrastructure.

The taxonomy reveals that ATLAS sits at the intersection of multiple research directions. Neighboring leaves include 'Task Execution Time and Resource Prediction' (four papers on ML models for job duration forecasting) and 'Cloud and Data Center Scheduling' (six papers on system implementations). The taxonomy's scope notes clarify that benchmark work should provide empirical infrastructure rather than novel algorithms or prediction techniques. ATLAS connects to these adjacent areas by offering a testbed for evaluating both prediction models and scheduling algorithms, bridging the gap between theoretical frameworks in 'Prediction Quality and Robustness' (four papers) and practical deployment in application domains.

Among twenty-four candidates examined via limited semantic search, none were found to clearly refute any of the three contributions. The ATLAS dataset contribution examined ten candidates with zero refutable matches; the LASched benchmark similarly examined ten candidates with no overlaps; the multi-stage ML model examined four candidates, also with no refutations. This suggests that within the scope of top-K semantic matches, the work occupies a relatively uncontested niche. However, the limited search scale means that more exhaustive exploration of adjacent fields—particularly production trace datasets in cloud computing or ML workload characterization—might reveal closer prior work not captured by this analysis.

Given the sparse benchmark leaf and the absence of refutations among examined candidates, the work appears to address a recognized gap in the field's empirical infrastructure. The limited search scope (twenty-four candidates) and the taxonomy's structure suggest that while the core contributions are distinct within learning-augmented scheduling, broader literature on workload traces or ML system benchmarks may contain related efforts not fully captured here. The analysis reflects what is visible through targeted semantic search rather than exhaustive coverage of all relevant domains.