Recriptografia Assistida por GPU em Servidores de Distribuição de Vídeo

Abstract

Proxy servers of video distribution systems must provide efficient memory management and also have to adopt protection mechanisms so that only allowed clients would have access to restricted video content. The intrinsic parallel nature of this problem plays a role on implementations that exploit parallel hardware available in off-the-shelf computing systems. With the spread of multicore GPUs, not only there is a need for empirical performance evaluation but also for other fundamental requirements such as energy consumption and total cost of implementation. In this work implementation alternatives of the recryptography module are compared, based on CUDA, Pthreads and OpenMP, using collected empirical results for different amounts of concurrent customers (given the cryptography throughput of 3,2 Mbps for a Fermi GPU and 1,8 Mbps for a multicore in a 1024 clients scenario), suggesting that the use of GPUs can reduce the total implementation cost of the module. Moreover, the present work proposes and validates, with satisfactory accuracy (ie with 9% and 14% diverging average for a given GPU of GT200 and Fermi families respectively), an adaptation of a mathematical model to forecast performance provided by different GPU architectures for the cryptography module implemented, which is cross-validated with empirical results collected for different architectures. To complete the feasibility analysis of a real implementation of the proposed module, this work intends to estimate the energy consumption of each experienced configuration by a GPU simulator where the results points to an average energy consumption about 60 W per kernel. Once known the main three metric (performance, energy consumption and price) it will be possible to determine the solution of better cost-benefit for the recryptography module.