Multimedia MPSoC (Multi-Processor System-on-Chip) has been widely used in today's consumer electronics. As many memory intensive IP (Intellectual Property) cores are integrated into a single chip, the performance of shared memory has been a critical issue. Conventional memory scheduling policies just schedule the existing requests in the queue and waste the potential scheduling space because they don't use the features of memory access modules. Existing policies often have to take a tradeoff between performance and QoS (Quality-of-Service). In this paper, we propose an exclusive memory scheduling policy by the observation of memory locality and latency tolerance of multimedia modules. We build a window mechanism and make IP cores monopolize the memory bus round-robin to get both performance and fairness. The experiment results show that the proposed scheduling policy improves the system bandwidth by 36.5% and reduces CPU latency by 29% under heaviest workload compared to a QoS-based control policy. In addition, our policy has low hardware cost and can cooperate with other schedulers.
展开▼
