Improving Efficiency of Embedded Multi-core Platforms with Scratchpad Memories

Audio/Video applications posses massive parallelism which can be exploited through todays multi and manycore parallel computing systems. However the challenge is to develop scalable models for such applications as the data intensive behavior prevents the application from scaling beyond a few cores. We introduce a novel idea to mitigate this problem by reducing the cache coherence traffic over the shared bus. The work includes a case study on the performance and scalability analysis of real-world audio & video applications followed by a technique to convert the existing multi-threaded applications into a more efficient, scalable and portable model using existing Scratchpad Memories (SPM). The results show significant improvement in performance, scalability and reduced power consumption, without increasing the programming and debugging effort.