Abstract
Frequency-Store is a novel columnar storage system designed specifically for image-based AI workloads. By treating image frequency components as individual columns, it allows applications to selectively fetch only necessary data and share it across multiple tasks. This approach addresses the significant data movement bottleneck in modern AI pipelines, achieving up to 11x speedup in inference and training while improving compression by 2.2x.
Introduction: The Data Movement Bottleneck
While GPU performance has scaled rapidly, the time required to move massive image datasets from storage to compute units has become a critical bottleneck. In many AI pipelines, over 95% of total inference time is consumed by data movement and decompression rather than actual model computation.
Traditional image formats like JPEG are monolithic; an application must download and decompress the entire file even if the AI model only requires low-frequency features for a prediction. Furthermore, when multiple applications process the same dataset, they often redundantly move the same data over the network, exacerbating the "data tax."
Frequency-Store Architecture
Frequency-Store applies the principles of relational column-stores to image data. It decomposes images into frequency components using the Discrete Cosine Transform (DCT) and organizes these components into shareable columns.
Batch-wise Columnar Storage
Instead of the traditional file-per-image paradigm, Frequency-Store groups images into batches (e.g., 256 images) and stores them column-by-column. This design reduces disk fragmentation and enables high-throughput bulk decompression, achieving ~40% faster reconstruction than traditional file-based methods.
Diagonal Column Abstraction
The system groups DCT coefficients along diagonals. Frequencies along the same diagonal share similar statistical distributions, which allows for optimized bit-encoding and efficient pruning. This "diagonal" approach provides a balance between fine-grained data access and system metadata overhead.
Key Insights
- Multimodal Data Characteristics: Different frequency components exhibit varying ranges and variances. Columnar storage allows each diagonal to utilize a custom bit-width, maximizing compression efficiency.
- Data Sharing: Frequency-Store defines a hierarchy of formats. A lightweight model might only read the first few diagonal columns, while a more complex model reads more. If both run concurrently, they share the common columns, eliminating redundant data movement.
Evaluation and Results
The system was evaluated across several datasets, including ImageNet subsets, medical imaging (Blood-cells), and weather prediction data.
- Performance: Frequency-Store achieves up to an 11x end-to-end speedup when the network is the bottleneck and 4.75x with local storage.
- Compression: By grouping similar values, it provides 2.2x better compression than state-of-the-art AI-specific storage solutions like Image Calculator.
- Scalability: Performance gains increase as more concurrent applications are added, thanks to the inherent data-sharing architecture.
Conclusion
Frequency-Store represents a shift toward "AI-native" image storage. By breaking images into shareable frequency columns and utilizing batch-wise storage, it effectively mitigates the data movement bottleneck that hinders large-scale AI deployment. Future work may extend this columnar approach to neural codecs and more complex spatial tasks like object detection or segmentation.
Summary updated