Performance Prediction Model for Block Ciphers on GPU Architectures

摘要

This paper presents a proposal of a performance prediction model of block ciphers on GPU architectures. The model comprises three phases: micro-benchmarks, analyzing code, and performance equations. Micro-benchmarks are developed in OpenCL considering scalability for GPU architectures of all kinds. Performance equations are developed, extracting some features of GPU architectures. Overall latencies of AES, Camellia, and SC2000, which covers all types of block ciphers, are inside the range of estimated latencies from the model. Moreover, assuming that out-of-order scheduling by Nvidia GPU works well, the model predicted overall encryption latencies respectively with 2.0 % and 8.8 % error for the best case on Nvidia Geforce GTX 580 and GTX 280. This model supports algebraic and bitslice implementation, although evaluation of the model is conducted in this paper only on table-based implementation.