FLEDA: Forecasting-Based Data Partition Placement for Efficient Deep Learning Recommendation Model Training
In recommendation and advertising systems, model training pipelines rely on high-throughput ingestion of fresh, collected behavioral data. These data are typically organized into time-partitioned, multi-level table paths. The access patterns driven by highly concurrent model training tasks exhibit complex cross-path correlations, multi-scale temporal dependencies, and frequent hot-cold transitions, making effective data placement under limited SSD capacity challenging. In this paper, we propose FLEDA, a forecasting-based data partition placement framework that selectively places daily-level training data partitions onto SSDs under constrained storage capacity. We first model historical data access patterns as multivariate and multiscale time series to capture both temporal dynamics and cross-path dependencies. By leveraging dependencies across two dimensions, we propose a model that extracts multi-scale cross-path temporal features and fuses mixture-of-experts to forecast the future throughput of all table paths. Besides, we design a ranking-based placement policy that balances predicted throughput, recent access history, and SSD residency continuity to maximize coverage of future data accesses while avoiding excessive data migration. Experiments on large-scale, real-world training workloads show that FLEDA consistently outperforms state-of-the-art multivariate time-series forecasting models and significantly improves SSDs' throughput coverage with low data migration overhead.