Can Software-based Transport Perform Like RDMA?

Abstract

RDMA delivers a high-performance data plane with minimal CPU overhead in datacenters, but its hardware-integrated control plane suffers from serious scalability limitations. Recent work on improving RoCEv2-based NIC (RNIC) scalability has helped, yet these solutions depend on specialized hardware features, limiting their practical adoption. In this paper, we ask the opposite question: can a software-based data plane match RDMA's performance while preserving the flexibility of a software control plane? We show that TCP performance can be dramatically improved by following the design philosophy of RoCE, which integrates kernel bypassing, segment offloading, and zero-copy buffer management. Our preliminary evaluation shows that a single CPU core can sustain over 1 Tbps on TCP transmission (TX) and 752 Gbps on reception (RX). These results demonstrate that software TCP stacks can achieve performance parity with modern RNICs, providing a flexible and practical alternative for high-performance networked systems.