TSO/GSO Segmentation offload
Letting the network card split a large data buffer into correctly sized packets, instead of the processor building each packet individually.
The cost in a network stack is not in the number of bytes but the number of packets: every frame means a trip through the layers, a header construction and a descriptor handed to the card. With segmentation offload the kernel hands over up to 64 kilobytes in one piece along with a header template, and the card cuts it into MSS-sized chunks, filling in sequence numbers and checksums itself. CPU load falls dramatically. GSO is the software variant that defers the split to the last possible moment inside the kernel, capturing part of the benefit even without hardware support. The receive-side equivalents are LRO and GRO. The side effect shows up when debugging: a packet capture shows implausibly large packets that were never on the wire.