Low Latency Implementations of CRYSTALS-Dilithium for Cloud Applications

This paper studies GPU-based implementations of CRYSTALS-Dilithium (Dilithium) algorithm, which is one of four post-quantum algorithms that NIST has standardized. We focus on the scenarios of batch processing and explore low latency implementations by parallelizing the critical components of Dilithium with the consideration of overhead due to inter-thread data sharing. We also investigate the techniques of using GPU memories and warp-level instructions for data sharing in GPU-based implementations of Dilithium. All of the efforts lead to effective low latency implementations that are suited to cloud applications. Finally, we conduct evaluations on two testbeds with NVIDIA GeForce RTX 3060 and RTX A4000, respectively. The experimental results show that our low latency implementations save up to 41 percent (resp. 33 percent) latency at low task loads on the former testbed (resp. the latter testbed) when compared with the baseline implementation. We also provide engineers with a practical strategy to achieve better overall performance for applications that the number of arrived tasks is not predictable.