Frequency-Store: Scaling Image AI by A Column-Store for Images

Utku Sirin, Aadit Saluja, Florian Klein, Jeremy Hsu, Stratos Idreos

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.

Inference time breakdown showing data movement as the primary bottleneck.

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.

Frequency-Store architecture for batch-wise columnar storage and data sharing across multiple applications.

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

Abstract

Artificial intelligence over images improves every aspect of modern human life and has shown great success across numerous applica- tions. However, it is costly to perform image AI. Image AI pipelines need to move heavy image files over the network so that many applications can concurrently process the images with varying re- source budgets and performance requirements. As a result, data movement dominates the end-to-end image AI cost. This work presents Frequency-Store, the first column-store for images. Our intuition is that images do not need to be consumed by image AI one whole image at a time. Instead, there are “components” of data within each image that can be consumed separately and thus also can be stored separately. This decomposition allows the sharing of data movement across image AI processing pipelines both for training and inference. Frequency-Store breaks images into columns and stores batches of images column by column rather than storing individual images file by file. It utilizes the inherent blocks and frequencies-based structure in image data and defines a novel column abstraction. Column-wise storage allows applications with various characteris- tics and resource demands to share data efficiently. Columns store data items with similar characteristics, allowing tight data represen- tations and efficient compression. We show that Frequency-Store improves inference/training time by up to 11x and compression ratio by up to 2.2x compared to state-of-the-art image AI stor
BibTeX
@inproceedings{sirin2025frequency,
  author = {Utku Sirin and Aadit Saluja and Florian Klein and Jeremy Hsu and Stratos Idreos},
  title = {{Frequency-Store: Scaling Image AI by A Column-Store for Images}},
  booktitle = {Cidr},
  year = {2025},
}
Figure