Performance and Energy Evaluation of Memory Organizations in NoC-Based MPSoCs under Latency and Task Migration

Abstract

This chapter presents a study on the performance and energy consumption arising from distinct memory organizations in an NoC-based MPSoC environment. This evaluation considers three sets of experiments. The first one evaluates the performance and energy efficiency of four different memory organizations in a situation where a single application is executed. In the second experiment, a traffic generator is responsible for the injection of synthetic traffic into the system, simulating the impact of the parallel execution of additional applications and increasing the latency of the NoC. Results show that, with a low NoC latency, the distributed memory presents better results for applications with low amount of data to be transferred. On the other hand, results suggest that shared and distributed shared memories present the best results for applications with high data transferring needs. In the second set of experiments, with higher NoC latency, for applications with low communication bandwidth requirements, a memory organization that is physically centralized and logically shared (called nDMA) is shown to have a smooth performance degradation when additional traffic rises up to 20% of the network capacity (22% degradation for an application demanding high communication, and 34% degradation for a low communication one). In contrast, a distributed memory model presents 2% of degradation in an application with high communication requirements, when traffic rises up to 20% of the network capacity, and reaches 19% of degradation in low communication ones. Shared and distributed shared memory models are shown to present lower tolerance to high latencies. A third set of experiments evaluates the performance of the four memory organization models in a situation of task migration, when a new application is launched and its tasks must be distributed among several nodes. Results show that the shared memory and distributed shared memory models have a better performance and energy savings than the distributed memory model in this situation. In addition, the nDMA memory model presents a smaller overhead when compared to the shared memory models and tends to reduce the traffic in the migration process due to the concentration of all memory modules in a single node of the network.