Acceleration of AES encryption on CUDA GPU

Abstract

GPU exhibits the capability for applications with a high level of parallelism despite its low cost. The support of integer and logical instructions by the latest generation of GPUs enables us to implement cipher algorithms more easily. However, decisions such as parallel processing granularity and memory allocation impose a heavy burden on programmers. Therefore, this paper presents results of several experiments that were conducted to elucidate the relation between memory allocation styles of variables of AES and granularity as the parallelism exploited from AES encoding processes using CUDA with an NVIDIA GeForce GTX285 (Nvidia Corp.). Results of these experiments showed that the 16 bytes/thread granularity had the highest performance. It achieved approximately 35 Gbps throughput. It also exhibited differences of memory allocation and granularity effects around 2%-30% for performance in standard implementation. It shows that the decision of granularity and memory allocation is the most important factor for effective processing in AES encryption on GPU. Moreover, implementation with overlapping between processing and data transfer yielded 22.5 Gbps throughput including the data transfer time.

HiCrypt: A Specialized Translator for Symmetric Block Cipher and GPGPU

“…The effect of the number of threads and file size are also shown. Iwai et al [17] further discussed granularity and the variety of memory allocations for variables of cipher programs. It showed that efficient CUDA programs made a difference of more than 50 times in performance.…”

“…It showed that efficient CUDA programs made a difference of more than 50 times in performance. Nishikawa et al [18], [19] applied the technique shown in [17] to other ciphers to validate the effectiveness of it. The result showed that implemented ciphers, Camellia, CIPHERUNICORN-A, Misty, and SC2000 also exhibit high-performance.…”

“…If the cache mechanism works well, then constant memory will provide the same low latency as local memory. However, the look-up table operation of block cipher requires random access [17]. Therefore, constant memory should be a wrong selection to store for such look-up tables.…”

“…Furthermore, allocating a variable for a key on the shared memory is valid. The key can be allocated on the shared memory or constant memory because each allocation does not make a difference for their performance [17].…”

Implementation of an XSL block cipher with MDS-matrix linear transformation on NVIDIA CUDA

“…In this paper we consider an XSL block cipher implementation over the finite field GF (2 8 ) with MDS-matrix linear transform and block-size 128 bit. We present different algorithmic optimization of this algorithm.…”