Acceleration of cryptographic
TCD-CS-2010-02.pdf (PDF) 2.667Mb
Graphics processing units (GPUs) can act as an attractive alternative to CPUs for general purpose computation in certain scenarios. Traditionally, the GPU has been developed to offload graphics processing from the CPU. In recent years the GPU has continued to become a more flexible and powerful device, responding to the demand of the games industry to execute more and more complex custom graphics algorithms. In the early 2000s a new approach to processing emerged, whereby non-graphics problems that suit a data parallel model could execute on the GPU at competitive or faster rates that the CPU. This performance gap has continued to grow, and as the GPU develops in terms of programming flexibility, the range of application spaces that benefit from GPU assistance widens. Adding to this trend, GPU vendors have started releasing programming frameworks specifically tailored to general purpose computation on GPUs. In light of these developments, there is intense research involving the use of GPUs for acceleration within many problem spaces. We advance the state of the art by presenting the capacity of the GPU to accelerate commonly used cryptographic functions. We investigate GPU acceleration of symmetric-key and asymmetric-key functions, fundamental components of modern cryptographic systems. We show that AES, a popular example of a symmetric-key function, can be competitive with the CPU on recent GPUs and outperform on contemporary GPUs. We illustrate the issues related to GPU support of symmetric-key modes of operations in various scenarios and present strategies for maintaining performance. We show that RSA, a popular example of an asymmetric-key function, can outperform the CPU when running on the GPU. For both symmetric-key and asymmetrickey approaches presented, not all cryptographic contexts suit the GPU and as such these contexts are highlighted. Also, both approaches are investigated for efficient batching of multiple requests within a single GPU call. Finally, the integration of GPU accelerated cryptography within an operating system abstraction layer and associated costs are presented.
Publisher:Trinity College Dublin. School of Computer Science and Statistics. Computer Science
Type of material:Thesis
Availability:Full text available