Frequency-domain-based performance analysis for the cell BE processor

Abstract

Frequency-domain analysis is routinely used to analyze complex DSP and communication systems, however its application in processor performance analysis has been quite limited. This paper explores the utility of frequency-domain analysis with respect to the Cell BE processor targeting streaming multimedia applications. The paper utilizes FFT as a means to characterize workloads. FFT is applied to either a program trace for complex applications, or to a ‘Z’ representation of key program kernels. The latter presentation utilizes the regular control-flow of multimedia kernels and achieves a concise representation of them. Such characterization identifies potential workload characteristics in terms of ‘repeated’ events. To predict performance, an analytical model of the Cell BE is constructed based on queuing networks. The parameters of the model are obtained, and reasoned about, using the frequency-domain's workload characterization. The performance model is compared with real experimental results using High Dynamic Range tone mapping operation, image processing kernel, and a DMA benchmark application.