Analysis and design optimization of communication process on message-passing-based MPSoC

By analyzing the message-passing communication process on MPSoC, efficient ways to enhance communication performance are concluded, which are reducing one-to-many message sending latency and improving efficiency of concurrently receiving multiple messages. A broadcast optimization strategy based on the hardware abstraction level to cut down the cost of data copy is proposed, which decreases one-to-many message sending latency efficiently. Fully taking into consideration reducing interaction times and improving the efficiency of long message communication when dealing with the bottleneck of concurrently receiving multiple messages, a novel network interface (NI) receiving strategy that combines a lookup table (LUT) mechanism and DMA mode is presented. Experimental results show that both broadcast optimization strategy and NI receiving strategy can achieve better performance. In a 64 × 64 matrix multiplication, the overall performance is increased approximately by 50%.