Catalog Home Page

Design of a scalable network interface to support enhanced TCP and UDP processing for high speed networks

Elbeshti, Mohamed (2014) Design of a scalable network interface to support enhanced TCP and UDP processing for high speed networks. PhD thesis, Murdoch University.

PDF - Front Pages
Download (480kB)
PDF - Whole Thesis
Download (5MB)


Communication networks have advanced rapidly in providing additional services, with improvements made to their bandwidth and the integration of advanced technology. As the speed of networks exceeds 10 Gbps, the time frame for completing the processing of TCP and UDP packets has become extremely short. The design and implementation of high performance Network Interfaces (NIs) that can support offload protocol functions for current and next-generation networks is challenging. In this thesis two software approaches are presented to enhance protocol processing of TCP and UDP in the network interface. A novel software Large Receive Offload (LRO) approach for enhancing the receiving side has been proposed. The LRO works by aggregating the incoming TCP and UDP packets into larger packets inside the NI’s buffer. The receiving side software has been improved to support out-of-order packets. The second proposed software solution is applied on the Large Send Offload (LSO). The proposed LSO function processing is implemented by segmenting TCP and UDP messages that are larger than the Maximum Transmission Unit to the Maximum Segment Size. New packet headers are generated for each new outgoing packet.

A scalable programmable NI based 32-bit RISC core is presented that can support 100 Gbps network speeds. Acceleration of the processing time frame required at the NI has been implemented to prevent hazards (such as Data Hazard and Control Hazard) during the execution of the LRO and the LSO functions. An R2000/3000 RISC has been used in order to test the LRO and LSO functions and to discover the instruction set that is most suitable. Following this the VHDL NI was implemented with three pipeline RISC cores, a simple DMA controller and Content Addressable Memory. An evaluation of the desired RISC clock rate that is required to process TCP and UDP streams at 100 Gbps was conducted. It was determined that a RISC core running at 752 MHz with a DMA clock of 3753 MHz was able to process packets 512 bytes or larger fast enough to support 100 Gbps network speeds.

Publication Type: Thesis (PhD)
Murdoch Affiliation: School of Engineering and Information Technology
Supervisor: Dixon, Michael and Koziniec, Terry
Item Control Page Item Control Page


Downloads per month over past year