Answers are grounded in this paper and cite their sources.
TCP pacing and TCP Segmentation Offload (TSO) are widely used mechanisms for improving network performance. Pacing, commonly implemented in software, spaces traffic over time to reduce bursts and thereby bottleneck queue length variation; TSO, commonly implemented in the network interface card (NIC), reduces CPU overhead by offloading packet segmentation. Because TSO happens after pacing, today's NICs emit microbursts that comprise dozens of segments for 1 Gbit/s and faster links. Disabling TSO can restore fine-grained pacing, but at the cost of the CPU efficiency that is required for high-rate TCP transmission. We demonstrate that TSO-induced microbursts are not an inherent cost of segmentation offload: modest NIC-side scheduling can retain TSO while restoring fine-grained pacing. For brevity, we refer to this as TSO pacing. It is commonly assumed that TSO pacing is not implemented on NICs because its memory requirements would raise NIC prices considerably. We propose a generic TSO pacing design that overcomes the memory challenge and present a specific implementation for Linux and a Netronome Agilio CX SmartNIC. The implementation changes only the NIC's driver and firmware and demonstrates that TSO can be combined with highly accurate pacing using limited additional NIC-side scheduling resources.
No takes yet. Share an insight, caveat, or question.
Mathisen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: