Reproduce “Dataflow-Oriented Classification and Performance Analysis of GPU-Accelerated Homomorphic Encryption” on OpenResearch

About this paper

Ai Nozaki, Takuya Kojima, Hideki Takase, Hiroshi Nakamura

Researchers at The University of Tokyo and Tsukuba University found that the most effective GPU acceleration strategies for CKKS homomorphic encryption operations vary by parameter configuration and GPU architecture, achieving up to 1.98x performance differences between approaches. Their work introduces a dataflow-oriented classification of optimizations to guide adaptive strategy selection.

Key insights

• The optimal GPU acceleration strategy for CKKS operations is highly dependent on both the specific CKKS parameter configuration and the target GPU architecture, necessitating adaptive optimization approaches.
• On modern consumer/workstation GPUs, the optimal strategy shifts from DigitParallel OutputBulk (DPOB) for smaller parameter sets to OutputChunked strategies (DPOC or DSOC) as the ciphertext memory footprint exceeds L2 cache capacity.
• High DRAM bandwidth, such as that in the NVIDIA A100, can significantly mitigate the performance impact of L2 cache misses, allowing strategies like DPOB, which leverage greater parallelism, to remain optimal across a wider range of parameter configurations.

Summary by alphaXiv.