Scalability of parallel video decoding on heterogeneous manycore architectures

Abstract

This paper presents an analysis of the scalability of the parallel video decoding on heterogeneous many core architectures. As benchmark, we use a highly parallel H.264/AVC video decoder that generates a large number of independent tasks. In order to translate task-level parallelism into performance gains both the video decoder and the architecture have been optimized. The video decoder was modified for exploiting coarse-grain frame-level parallelism in the entropy decoding kernel which has been considered the main bottleneck. Second, a heterogeneous combination of cores is evaluated for executing different type of tasks. Finally, an evaluation of the memory requirements of the whole system has been carried out. Experiments conducted using a trace-driven simulation methodology shows that the evaluated system exhibits a good parallel scalability up to 68 cores. At this point the parallel video decoder is able to decode more than 200 HD frames per second using simple low power processors.