Scalable Image AI via Self-Designing Storage

Utku Sirin, Victoria Kauffman, Aadit Saluja, Florian Klein, Jeremy Hsu, Vlad Cainamisir, Qitong Wang, Konstantinos Kopsinis, Stratos Idreos

Abstract:

Image AI has the potential to improve every aspect of human life. Image AI, however, is very ex-
pensive. We identify that the root cause of the problem is a long-overlooked and largely unexplored
dimension: storage. Most images today are stored as JPEG files. JPEG is designed for digital photog-
raphy. It maximally compresses images with minimal loss in visual quality. We observe that JPEG is a
fixed design. AI problems, however, are diverse; every problem is unique in terms of how data should
be stored and processed. Using a fixed design, such as JPEG, for all problems results in excessive data
movements and costly image AI systems. This paper presents Image Calculator, a self-designing storage
system that finds the optimal storage for a given image AI task. The Image Calculator achieves this by
identifying design primitives for image storage and creating a design space comprising thousands of
storage formats based on these design primitives, each capable of storing and representing data differ-
ently, with varying accuracy, inference and training time, as well as space consumption trade-offs. It
efficiently searches within this design space by building performance models and using locality among
its storage formats. It exploits the inherent frequency structure in image data to efficiently serve infer-
ence and training requests. We evaluate the Image Calculator across a diverse set of datasets, tasks,
models, and hardware. We show that Image Calculator can generate storage formats that reduce end-
to-end inference and training times by up to 14.2x and consumed space by up to 8.2x with little or
no loss in accuracy, compared to state-of-the-art image storage formats. Its incremental computation
and data-sharing schemes over frequency components allow scalable inference- and training-serving
systems.