Deferred segmentation for wire-speed transmission of large TCP frames over standard GbE networks

Abstract

"Deferred Segmentation" (DS) is a novel approach to enhance the TCP transmission rate of large frames over standard GbE networks. DS allows large TCP frames through the senders's TCP/IP stack, and the NIC breaks them down into TCP segments of standard Ethernet MTU size. DS doubles TCP/IP performance, pushing the TCP transmission rate to the wire speed and/or reducing the host CPU load. Unlike when the protocol stack is split between the host CPU and the NIC, or is offloaded to the NIC, DS requires, at most, a small NIC development effort and no modification of legacy applications or OS TCP/IP stacks. DS need only be implemented in the sender and is transparent to the receivers. If traffic is mostly outbound and is split among multiple receivers, so doing will not reduce performance. Finally, DS can even be implemented entirely in the NIC's device driver and still improve performance.