The Future of Fully Homomorphic Encryption System: from a Storage I/O Perspective

TL;DR

This paper investigates storage I/O as a practical bottleneck for deploying Fully Homomorphic Encryption (FHE) in cloud environments. Using CKKS-based workloads (ResNet-20 inference and HELR training) and two accelerators (ASIC Sharp and GPU TensorFHE), it quantifies how data movement from external storage to on-chip memory drastically reduces performance, even when computation is accelerated. The study shows that I/O overhead can cause up to hundreds of times slowdown, with performance depending on locality, distribution, and the specific FHE parameters and operators used. The findings highlight storage I/O as a critical factor in real-world FHE deployments and propose directions such as near-data processing and locality-aware scheduling to bridge the gap toward field deployment.

Abstract

Fully Homomorphic Encryption (FHE) allows computations to be performed on encrypted data, significantly enhancing user privacy. However, the I/O challenges associated with deploying FHE applications remains understudied. We analyze the impact of storage I/O on the performance of FHE applications and summarize key lessons from the status quo. Key results include that storage I/O can degrade the performance of ASICs by as much as 357$\times$ and reduce GPUs performance by up to 22$\times$.

Figures (7)

• Figure 1: Speedup of FHE accelerators.
• Figure 2: FHE workflow.
• Figure 3: Experiment setup.
• Figure 4: Performance under different storage: (a) ASIC, (b) GPU. The ASIC's baseline performance of ResNet-20 and HELR is 99ms and 2.5ms, and GPU's baseline performance is 4.9s and 220ms. The I/O overhead significantly decrease the performance of ASIC and GPU.
• Figure 5: Performance under varying cache hit ratio: (a) Resnet-20 (b) HELR. Different storages and different applications need different cache hit ratios to achieve a fixed target performance.