Data-oriented performance analysis of SHA-3 candidates on FPGA accelerated computers

The SHA-3 competition organized by NIST has triggered significant efforts in performance evaluation of cryptographic hardware and software. These benchmarks are used to compare the implementation efficiency of competing hash candidates. However, such benchmarks test the algorithm in an ideal setting, and they ignore the effects of system integration. In this contribution, we analyze the performance of hash candidates on a high-end computing platform consisting of a multi-core Xeon processor with an FPGA-based hardware accelerator. We implement two hash candidates, Keccak and SIMD, in various configurations of multi-core hardware and multi-core software. Next, we vary application parameters such as message length, message multiplicity, and message source. We show that, depending on the application parameter set, the overall system performance is limited by three possible performance bottlenecks, including limitations in computation speed, in communication band-width, and in buffer storage. Our key result is to demonstrate the dependency of these bottlenecks on the application parameters. We conclude that, to make sound system design decisions, selecting the right hash candidate is only half of the solution: one must also understand the nature of the data stream which is hashed.