Protocol conversions for the Aethereal Networks-on-Chip

Abstract

Bus based interconnects are commonly used to connect Intellectual Properties (IPs) on System-on-Chip (SoC). However, as the number of high performance IPs with large communication requirements in a Multi Processor SoC (MPSoC) increases, the bus interconnects become a communication bottleneck. To overcome this limitation, the bus based interconnects are replaced by Networks-on-Chip (NoC) as the interconnect for IPs in an MPSoC. The ability to support multiple protocols, both legacy and newer, is essential for leveraging the advantages of the NoC. This thesis describes the protocol conversions, or shells, we design for the aethereal NoC, to provide a unified message format for the PLB, OPB and FSL protocols. Furthermore, to tolerate the latencies when accessing memory and to increase the throughput when possible, we add support for posted write and prefetch read in the shells and we design a mechanism to coalesce multiple single transactions into burst transaction when possible. To validate our design, we prototype it on a Virtex-II PRO FPGA. We use both synthetic applications as well as real life applications to benchmark and analyze the effect of factors, like computation-to-communication ratio, various link bandwidths and compiler loop unrolling on the performance of the system. The results show that reducing the link bandwidth up to a certain point does not affect the performance anymore, as the latency associated with NoC internals becomes dominant. Also the burst transaction can sustain the performance of the system up to a certain point, when the link bandwidth is reduced. The best result is obtained using a shell with posted write and prefetch read.