Considering Barrier Synchronization Overhead in Parallelizing Cryptographic Algorithms

Abstract

As a multi-core processor has been announced recently, many multimedia applications are being parallelized in order to reduce the execution time. Although the size of multimedia data to be protected is huge, it is difficult to efficiently speed up some cryptography algorithms due to data dependence. In this paper, we propose a general approach which parallelizes cryptographic algorithms having data dependency in order to reduce the execution time for huge data. After analyzing the parallel overhead with AES as an example of cryptographic algorithms having data dependency, we found the barrier synchronization overhead caused by data dependency is the most important one with some multi-core processors. Then, we overlap computations in order to solve the limited parallelism inherent in AES. That is, we interpret the problem as an application-level pipeline and derive the optimum number of pipeline stages by considering the barrier synchronization overhead.