Accelerating RSA with Fine-Grained Parallelism Using GPU

RSA is a public key cryptography widely used for end-to-end authentication and key exchange in various Internet protocols, such as SSL and TLS. Compared with symmetric cryptography, the cryptographic operations in RSA is much more time consuming. This brings pressure on performance to service providers using secure protocols, and hinders these protocols from being more widely used. Graphics Processing Units (GPUs) are increasingly used for intensive data parallelism general purpose computing. GPUs often provide better throughput than CPUs at the same cost. In this paper, we propose a new approach to parallelize Montgomery multiplication under the Single Instruction Multiple Thread (SIMT) threading model of GPUs, and construct a parallel RSA implementation based on this approach, combining with other optimization techniques both in the algorithmic level and implementation level. The performance evaluation shows our RSA implementation achieves a record-breaking latency for RSA decryption implementations on GPUs: 2.6 ms for RSA-1024 and 6.5 ms for RSA-2048. The peak throughtput of decryptions per second of our implementation reaches 5,244 for RSA-2048 and 34,981 for RSA-1024 respectively, which is much faster than existing integer-based implementations. The peak throughput of our implementation is slightly slower than the fastest floating-point based implementation, while the latency of our implementation is 3 times faster.