Design and implementation of hardware-based low latency TCP offload engine for 10 Gbps Ethernet

With the data growth related to the increase of devices with network access, the high speed Ethernet is widely used in various fields from big data center to high frequency transaction (HFT). However, the conventional software-based protocol stack consumes large amount of CPU time at full transmission rate, and leads to low performance such as high latency and low throughput. This paper presents a hardware based TCP offload engine (TOE) operating at full 10 Gbps throughput full-duplex with the low latency. Features supporting high speed network such as congestion avoidance and timestamp are implemented. The proposed TOE is verified by interacting with the software protocol stack. The performance measurement shows our TOE can provide transmission rate up to 10 Gbps as both transmitter and receiver. The input-to-output latency for 100 bytes payload and 64 bytes header with timestamp is about 90 nanoseconds compared with the 140 nanoseconds implemented by Dini Group.