Abstract:

The current development of the Ethernet technology is much faster than that of memory and CPU technologies, and the memory access and CPU processing network stack have become the bottleneck of TCP performance on end systems. The constantly increasing network bandwidth burdens  the CPU severely, and approximately 1GHz CPU resource is needed to process 1Gbps network traffic. We therefore take a multicore NPU as the NIC and the TCP's checksum verification and packets reordering functions are offloaded. Small TCP packets aggregated into fewer but larger packets by the multicore NPU, thus reducing both the number of packets processed by network stack and the number of interrupts generated by the NIC, and eventually improving the TCP performance on end systems. Experimental results show that 4.9 Gbps TCP receive data throughput can be achieved in a 10Gbps network.

Key words: TCP packets reordering;TCP data receive offload;LRO;TOE;multicore NPU