Maximizing GPU Parallelism for a High-performance Cryptanalysis System

Modern cryptanalysis techniques utilize GPUs, owing to their parallel processing power, to significantly improve the speed of cryptanalysis tasks. Thus, recent research directions have focused on maximizing parallelism to complete large-scale cryptanalysis in a short time. For this, previous works utilized NVIDIA’s technologies such as Hyper-Q, Multi-Process Service (MPS), and Multi-Instance GPU (MIG) for their cryptanalysis system. However, no effort has been made to evaluate whether integrating these technologies improves parallelism compared to a single one. In this paper, we design and implement a cryptanalysis system that integrates MPS and MIG. To evaluate this system, we develop three versions of Strassen’s matrix multiplication programs for large-scale cryptanalysis. We evaluate the performance of these programs using CUDA, cuBLAS, and a combination of cuBLAS and MPI, respectively. Our results show that with both MPS and MIG enabled, the computation time for matrix multiplication is approximately 1.42×, 1.28×, and 2.02× faster.