Abstract: Systems, apparatusses, and methods are disclosed for transmission control protocol (TCP) segmentation offload (TSO). A hardware TSO engine is capable of handling segmentation of data packets and consequent header field mutation of hundreds of flows simultaneously. The TSO engine generates data pointers in order to “cut up” the payload data of a data packet, thereby creating multiple TCP segments. Once the data of the data packet has been fetched, the TSO engine “packs” the potentially-scattered chunks of data into TCP segments, and recalculates each TCP segment's internet protocol (IP) length, IP identification (ID), IP checksum, TCP sequence number, and TCP checksum, as well as modifies the TCP flags. The TSO engine is able to rapidly switch contexts, and share the control logic amongst all flows.

Abstract: The disclosed packet scheduler implements the deficit round robin (DRR) approximation of weighted fair queuing (WFQ), and is capable of achieving complete fairness across several hundred source flows, for example, each of which can be mapped to one of several destination ports. In addition to achieving fairness, the packet scheduler allows the user to map one or more optional strict-priority flows to each port. The packet scheduler keeps these strict-priority flows “outside” of the group of flows for which fairness is enforced. Each destination port can be optionally configured to chop its data packets into sub-packet pieces. The packet scheduler works in two mutually orthogonal dimensions: (1.) it selects destination ports based on a round-robin scheme, or using another method, such as guaranteed rate port scheduling (GRPS), and (2.) it implements optional strict-priority scheduling, and DRR scheduling.

Abstract: A guaranteed rate port scheduler (GRPS) is used for serving multiple destination ports simultaneously without under-runs, even if the total bandwidth of the ports is more than the bandwidth capability of the device. Certain network protocols, such as Ethernet, do not allow “gaps” (called under-runs) to occur between bits of a packet on the wire. If a network device is transmitting packets to several such ports at the same time and the combined bandwidth of these ports is more than the device can source, under-runs begin to occur within the transmitted packets. The disclosed GRPS solves this problem by: (a) the GRPS serves only as many destination ports at a given time as can be “handled”, and (b) the GRPS fairly selects new destination ports to serve after every end-of-frame data packet transmission by effectively “de-rating” the statistical bandwidth of each destination port in proportion to the diminished capacity of the device.