Scalable Multi-Session TCP Offload Engine for Latency-Sensitive Applications

Abstract

Latency-sensitive applications, such as Network File System (NFS) and High-Frequency Trading (HFT), demand ultra-low latency in network communications. These applications usually need more than one TCP session to guarantee Quality of Service (QoS) in case of communication interruption. This paper introduces a scalable multi-session TCP Offload Engine (TOE) for latency-sensitive applications which reduces the delay using the kernel bypass approach. The input-output receiving latency of a 48-byte-payload packet is 262.4 ns, and the sending latency of the same size packet is 179.3 ns. The latencies grow linearly with the amount of data at the rate of 12.8 ns per 8 bytes. The latencies are irrelevant to the TCP session number, which shows the scalability of our implementation.